[Attention] MLA - cutlass decode with unresticted num_heads

.buildkite/scripts/hardware_ci/run-hpu-test.sh

@@ -6,17 +6,19 @@ set -exuo pipefail
 
 # Try building the docker image
 cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
-FROM gaudi-base-image:latest
+FROM 1.22-413-pt2.7.1:latest
 
 COPY ./ /workspace/vllm
 
 WORKDIR /workspace/vllm
 
+RUN pip install -v -r requirements/hpu.txt
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
 ENV no_proxy=localhost,127.0.0.1
 ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
 
-RUN VLLM_TARGET_DEVICE=empty pip install .
-RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
 
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils

.buildkite/test-pipeline.yaml

@@ -117,7 +117,7 @@ steps:
   commands:
   - pytest -v -s core
 
-- label: Entrypoints Test (LLM) # 40min
+- label: Entrypoints Test # 40min
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -125,6 +125,8 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -133,21 +135,9 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
   - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
-
-- label: Entrypoints Test (API Server) # 40min
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/tests"
-  fast_check: true
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   mirror_hardwares: [amdexperimental]
@@ -640,18 +630,6 @@ steps:
     # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
     # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*
 
-- label: Transformers Nightly Models Test
-  working_dir: "/vllm-workspace/"
-  optional: true
-  commands:
-    - pip install --upgrade git+https://github.com/huggingface/transformers
-    - pytest -v -s tests/models/test_initialization.py
-    - pytest -v -s tests/models/multimodal/processing/
-    - pytest -v -s tests/models/multimodal/test_mapping.py
-    - python3 examples/offline_inference/basic/chat.py
-    - python3 examples/offline_inference/audio_language.py --model-type whisper
-    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
-
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 

.gemini/config.yaml

@@ -1,6 +0,0 @@
-# https://developers.google.com/gemini-code-assist/docs/customize-gemini-behavior-github
-have_fun: false  # Just review the code
-code_review:
-  comment_severity_threshold: HIGH  # Reduce quantity of comments
-  pull_request_opened:
-    summary: false  # Don't summarize the PR in a separate comment

.github/CODEOWNERS

@@ -16,7 +16,6 @@
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm
 /vllm/entrypoints @aarnphm
-/vllm/compilation @zou3519 @youkaichao
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.34.0
+  rev: v1.32.0
   hooks:
   - id: typos
 - repo: https://github.com/PyCQA/isort
@@ -166,7 +166,7 @@ repos:
     language: python
     types: [python]
     pass_filenames: true
-    files: vllm/config.py|tests/test_config.py|vllm/entrypoints/openai/cli_args.py
+    files: vllm/config.py|tests/test_config.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -171,6 +171,16 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
+#
+# Set nvcc fatbin compression.
+#
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+    list(APPEND VLLM_GPU_FLAGS "-Xfatbin" "-compress-all" "-compress-mode=size")
+  endif()
+endif()
+
+
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
 # setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.

README.md

@@ -63,6 +63,8 @@ vLLM is fast with:
 - Speculative decoding
 - Chunked prefill
 
+**Performance benchmark**: We include a performance benchmark at the end of [our blog post](https://blog.vllm.ai/2024/09/05/perf-update.html). It compares the performance of vLLM against other LLM serving engines ([TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [SGLang](https://github.com/sgl-project/sglang) and [LMDeploy](https://github.com/InternLM/lmdeploy)). The implementation is under [nightly-benchmarks folder](.buildkite/nightly-benchmarks/) and you can [reproduce](https://github.com/vllm-project/vllm/issues/8176) this benchmark using our one-click runnable script.
+
 vLLM is flexible and easy to use with:
 
 - Seamless integration with popular Hugging Face models

csrc/attention/attention_kernels.cuh

@@ -24,7 +24,6 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
-#include "cuda_compat.h"
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -34,6 +33,12 @@ typedef __hip_bfloat16 __nv_bfloat16;
   #include "../quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
+
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
@@ -665,6 +670,7 @@ __global__ void paged_attention_v2_reduce_kernel(
 
 }  // namespace vllm
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/attention/mla/cutlass_sm100_mla/device/sm100_mla.hpp

@@ -125,24 +125,16 @@ public:
   }
 
   static void set_split_kv (KernelArguments& args) {
-    // printf("set_split_kv start");
     if (args.split_kv >= 1) return;
     auto [H, K, D, B] = args.problem_shape;
-    // std::cout << H << " " << K << " " << D << " " << B << "\n";      
     int sm_count = args.hw_info.sm_count;
-    // printf("    sm_count = %d\n", sm_count);
     int max_splits = ceil_div(K, 128);
-    max_splits = min(16, max_splits);
-    // printf("    max_splits = %d\n", max_splits);
     int sms_per_batch = max(1, sm_count / B);
-    // printf("    sms_per_batch = %d\n", sms_per_batch);
     int split_heur = min(max_splits, sms_per_batch);
     int waves = ceil_div(B * split_heur, sm_count);
     int k_waves = ceil_div(max_splits, split_heur);
     int split_wave_aware = ceil_div(max_splits, k_waves);
     args.split_kv = split_wave_aware;
-    // printf("    args.split_kv = %d\n", args.split_kv);
-
   }
 
   /// Determines whether the GEMM can execute the given problem.

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_fmha_mla_tma_warpspecialized.hpp

@@ -1446,7 +1446,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
     TiledMmaPV tiled_mma_pv;
 
     Tensor tStS =  partition_fragment_C(tiled_mma_qk, select<0,1>(TileShapeQK{}));
-    Tensor tItI =  partition_fragment_C(tiled_mma_pv, select<0,1>(TileShapePV{}));
+    Tensor tOtO =  partition_fragment_C(tiled_mma_pv, select<0,1>(TileShapePV{}));
 
     tiled_mma_pv.accumulate_ = UMMA::ScaleOut::Zero;
 
@@ -1519,7 +1519,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
 
           int read_stage = pipeline_load_pv_consumer_state.index();
 
-          tItI.data() = uint32_t(TmemAllocation::kO0) + j * uint32_t(TmemAllocation::kSizeAccO);
+          tOtO.data() = uint32_t(TmemAllocation::kO0) + j * uint32_t(TmemAllocation::kSizeAccO);
           tiled_mma_pv.accumulate_ = acc_flag;
 
           CUTLASS_PRAGMA_UNROLL
@@ -1527,7 +1527,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
             cute::gemm(tiled_mma_pv,
                        tOrP(_,_,k_block, make_coord(i, pipeline_p_mma_consumer_state.index())),
                        tOrVC(_,_,k_block,read_stage),
-                       tItI);
+                       tOtO);
             tiled_mma_pv.accumulate_ = UMMA::ScaleOut::One;
           }
 
@@ -1554,7 +1554,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
 
         int read_stage = pipeline_load_pv_consumer_state.index();
 
-        tItI.data() = uint32_t(TmemAllocation::kO0) + j * uint32_t(TmemAllocation::kSizeAccO);
+        tOtO.data() = uint32_t(TmemAllocation::kO0) + j * uint32_t(TmemAllocation::kSizeAccO);
         tiled_mma_pv.accumulate_ = acc_flag;
 
         CUTLASS_PRAGMA_UNROLL
@@ -1562,7 +1562,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           cute::gemm(tiled_mma_pv,
                      tOrP(_,_,k_block, make_coord(i, pipeline_p_mma_consumer_state.index())),
                      tOrVC(_,_,k_block,read_stage),
-                     tItI);
+                     tOtO);
           tiled_mma_pv.accumulate_ = UMMA::ScaleOut::One;
         }
 
@@ -1716,12 +1716,12 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
 
     TiledMmaPV tiled_mma_pv;
 
-    Tensor tItI = partition_fragment_C(tiled_mma_pv, select<0,1>(TileShapePV{}));
-    tItI.data() = tmem_o;
+    Tensor tOtO = partition_fragment_C(tiled_mma_pv, select<0,1>(TileShapePV{}));
+    tOtO.data() = tmem_o;
 
-    CUTE_STATIC_ASSERT_V(shape<1>(tItI) == _1{});
-    CUTE_STATIC_ASSERT_V(shape<2>(tItI) == _1{});
-    Tensor tAcc = tItI(make_coord(_,_),_0{},_0{});
+    CUTE_STATIC_ASSERT_V(shape<1>(tOtO) == _1{});
+    CUTE_STATIC_ASSERT_V(shape<2>(tOtO) == _1{});
+    Tensor tAcc = tOtO(make_coord(_,_),_0{},_0{});
 
     auto cta_tiler_pv = take<0,2>(typename CollectiveMmaPV::CtaShape_MNK{});
     Tensor gO = make_tensor(make_gmem_ptr((ElementAcc*) nullptr), cta_tiler_pv, make_stride(0, 0));
@@ -1773,12 +1773,12 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
 
     TiledMmaPV tiled_mma_pv;
 
-    Tensor tItI = TiledMmaPV::make_fragment_C(partition_shape_C(TiledMmaPV{}, take<0, 2>(TileShapePV{})));
-    tItI.data() = tmem_o;
+    Tensor tOtO = TiledMmaPV::make_fragment_C(partition_shape_C(TiledMmaPV{}, take<0, 2>(TileShapePV{})));
+    tOtO.data() = tmem_o;
 
-    CUTE_STATIC_ASSERT_V(shape<1>(tItI) == _1{});
-    CUTE_STATIC_ASSERT_V(shape<2>(tItI) == _1{});
-    Tensor tAcc = tItI(make_coord(_,_),_0{},_0{});
+    CUTE_STATIC_ASSERT_V(shape<1>(tOtO) == _1{});
+    CUTE_STATIC_ASSERT_V(shape<2>(tOtO) == _1{});
+    Tensor tAcc = tOtO(make_coord(_,_),_0{},_0{});
 
     auto [H, K, D, B] = problem_shape;
     auto [D_latent, D_rope] = D;

csrc/attention/mla/sm100_cutlass_mla_kernel.cu

@@ -18,7 +18,6 @@ limitations under the License.
  * Taken from SGLANG PR https://github.com/sgl-project/sglang/pull/6929
  * by Alcanderian JieXin Liang
  */
-#include "core/registration.h"
 
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
@@ -271,13 +270,4 @@ int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_ba
 }
 
 #endif
-
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
-  m.impl("sm100_cutlass_mla_decode", &sm100_cutlass_mla_decode);
-}
-
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CatchAll, m) {
-  m.impl("sm100_cutlass_mla_get_workspace_size", &sm100_cutlass_mla_get_workspace_size);
-}
-
 // clang-format on

csrc/attention/paged_attention_v1.cu

@@ -18,7 +18,12 @@
  */
 
 #include "attention_kernels.cuh"
-#include "cuda_compat.h"
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -182,6 +187,7 @@ void paged_attention_v1(
                              CALL_V1_LAUNCHER_BLOCK_SIZE)
 }
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/attention/paged_attention_v2.cu

@@ -18,7 +18,12 @@
  */
 
 #include "attention_kernels.cuh"
-#include "cuda_compat.h"
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -192,6 +197,7 @@ void paged_attention_v2(
                              CALL_V2_LAUNCHER_BLOCK_SIZE)
 }
 
+#undef WARP_SIZE
 #undef MAX
 #undef MIN
 #undef DIVIDE_ROUND_UP

csrc/cpu/sgl-kernels/common.h

@@ -58,7 +58,7 @@ namespace {
 
 #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
 #define CHECK_LAST_DIM_CONTIGUOUS(x) \
-  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimension")
+  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimention")
 
 #define CHECK_INPUT(x) \
   CHECK_CPU(x);        \

csrc/cpu/sgl-kernels/gemm.h

@@ -126,7 +126,7 @@ void fused_experts_int4_w4a16_kernel_impl(
     int64_t topk,
     int64_t num_tokens_post_pad);
 
-// shared expert implementation for int8 w8a8
+// shared expert implememntation for int8 w8a8
 template <typename scalar_t>
 void shared_expert_int8_kernel_impl(
     scalar_t* __restrict__ output,

csrc/cpu/sgl-kernels/gemm_int8.cpp

@@ -41,7 +41,7 @@ struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
     __m512  vd0;
     __m512  vd1[COLS];
 
-    // oops! 4x4 spills but luckily we use 4x2
+    // oops! 4x4 spills but luckly we use 4x2
     __m512 vbias[COLS];
 
     // [NOTE]: s8s8 igemm compensation in avx512-vnni

csrc/cpu/sgl-kernels/vec.h

@@ -37,7 +37,7 @@ inline Vectorized<at::BFloat16> convert_from_float_ext<at::BFloat16>(const Vecto
 #define CVT_FP16_TO_FP32(a) \
     _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
 
-// this doesn't handle NaN.
+// this doesn't hanel NaN.
 inline __m512bh cvt_e4m3_bf16_intrinsic_no_nan(__m256i fp8_vec) {
   const __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
 

csrc/cuda_compat.h

@@ -4,10 +4,10 @@
   #include <hip/hip_runtime.h>
 #endif
 
-#if defined(USE_ROCM) && defined(__GFX9__)
-  #define WARP_SIZE 64
-#else
+#ifndef USE_ROCM
   #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
 #endif
 
 #ifndef USE_ROCM

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -7,11 +7,7 @@
 
 #include <c10/util/BFloat16.h>
 #include <c10/util/Half.h>
-#ifdef USE_ROCM
-    #include <c10/hip/HIPException.h>  // For C10_HIP_CHECK and C10_HIP_KERNEL_LAUNCH_CHECK
-#else
-    #include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
-#endif
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
 
 #ifndef USE_ROCM
     #include <cub/block/block_load.cuh>
@@ -316,25 +312,19 @@ void selective_scan_fwd_launch(SSMParamsBase &params, cudaStream_t stream) {
     // kIsVariableB, kIsVariableC and kHasZ are all set to True to reduce binary size
     constexpr bool kIsVariableB = true;
     constexpr bool kIsVariableC = true;
+    constexpr bool kHasZ = true;
     BOOL_SWITCH(params.seqlen % (kNThreads * kNItems) == 0, kIsEvenLen, [&] {
-        BOOL_SWITCH(params.z_ptr != nullptr , kHasZ, [&] {
-            BOOL_SWITCH(params.query_start_loc_ptr != nullptr , kVarlen, [&] {
-                using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t>;
-                constexpr int kSmemSize = Ktraits::kSmemSize + kNRows * MAX_DSTATE * sizeof(typename Ktraits::scan_t);
-                dim3 grid(params.batch, params.dim / kNRows);
-                auto kernel = &selective_scan_fwd_kernel<Ktraits>;
-                if (kSmemSize >= 48 * 1024) {
-#ifdef USE_ROCM
-                    C10_HIP_CHECK(hipFuncSetAttribute(
-                        reinterpret_cast<const void*>(kernel), hipFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-#else
-                    C10_CUDA_CHECK(cudaFuncSetAttribute(
-                        kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
-#endif
-                }
-                kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
-                C10_CUDA_KERNEL_LAUNCH_CHECK();
-            });
+        BOOL_SWITCH(params.query_start_loc_ptr != nullptr , kVarlen, [&] {
+            using Ktraits = Selective_Scan_fwd_kernel_traits<kNThreads, kNItems, kNRows, kIsEvenLen, kIsVariableB, kIsVariableC, kHasZ,  kVarlen, input_t, weight_t>;
+            constexpr int kSmemSize = Ktraits::kSmemSize + kNRows * MAX_DSTATE * sizeof(typename Ktraits::scan_t);
+            dim3 grid(params.batch, params.dim / kNRows);
+            auto kernel = &selective_scan_fwd_kernel<Ktraits>;
+            if (kSmemSize >= 48 * 1024) {
+                C10_CUDA_CHECK(cudaFuncSetAttribute(
+                    (void *) kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+            }
+            kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+            C10_CUDA_KERNEL_LAUNCH_CHECK();
         });
     });
 }
@@ -622,20 +612,19 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
 
     at::Tensor z, out_z;
     const bool has_z = z_.has_value();
-    if (has_z) {
-        z = z_.value();
-        TORCH_CHECK(z.scalar_type() == input_type);
-        TORCH_CHECK(z.is_cuda());
-        TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
-        if (varlen){
-            CHECK_SHAPE(z, dim, seqlen);
-        } else {
-            CHECK_SHAPE(z, batch_size, dim, seqlen);
-        }
-        
-        out_z = z;
+    TORCH_CHECK(has_z, "has_z = False is disabled in favor of reduced binary size")
+    z = z_.value();
+    TORCH_CHECK(z.scalar_type() == input_type);
+    TORCH_CHECK(z.is_cuda());
+    TORCH_CHECK(z.stride(-1) == 1 || z.size(-1) == 1);
+    if (varlen){
+        CHECK_SHAPE(z, dim, seqlen);
+    } else {
+        CHECK_SHAPE(z, batch_size, dim, seqlen);
     }
 
+    out_z = z;
+
     // Right now u has BHL layout and delta has HBL layout, and we want out to have HBL layout
     at::Tensor out = delta;
     TORCH_CHECK(ssm_states.scalar_type() == input_type);
@@ -664,3 +653,4 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
         selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
     });
 }
+

csrc/ops.h

@@ -167,6 +167,19 @@ void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
                         torch::Tensor const& seq_lens,
                         torch::Tensor const& page_table, double scale);
 
+void sm100_cutlass_mla_decode(
+    torch::Tensor const& out, torch::Tensor const& q_nope,
+    torch::Tensor const& q_pe, torch::Tensor const& kv_c_and_k_pe_cache,
+    torch::Tensor const& seq_lens, torch::Tensor const& page_table,
+    torch::Tensor const& workspace, double sm_scale,
+    int64_t num_kv_splits =
+        1 /* Set to 1 to avoid cuda_graph issue by default. */);
+
+int64_t sm100_cutlass_mla_get_workspace_size(
+    int64_t max_seq_len, int64_t num_batches, int64_t sm_count = 0,
+    int64_t num_kv_splits =
+        1 /* Set to 1 to avoid cuda_graph issue by default. */);
+
 torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
 
 #ifndef USE_ROCM

csrc/torch_bindings.cpp

@@ -521,14 +521,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                         Tensor page_table, Tensor workspace, float "
       "scale,"
       "                         int num_kv_splits) -> ()");
-  // conditionally compiled so impl in source file
+  ops.impl("sm100_cutlass_mla_decode", torch::kCUDA, &sm100_cutlass_mla_decode);
 
   // SM100 CUTLASS MLA workspace
   ops.def(
       "sm100_cutlass_mla_get_workspace_size(int max_seq_len, int num_batches,"
       "                                     int sm_count, int num_kv_splits) "
       "-> int");
-  // conditionally compiled so impl in source file
+  ops.impl("sm100_cutlass_mla_get_workspace_size",
+           &sm100_cutlass_mla_get_workspace_size);
 
   // Compute NVFP4 block quantized tensor.
   ops.def(

docker/Dockerfile

@@ -63,7 +63,7 @@ ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL=https://download.pytorch.org/whl/nightly
 ARG PIP_KEYRING_PROVIDER=disabled
 ARG UV_KEYRING_PROVIDER=${PIP_KEYRING_PROVIDER}
 
-# Flag enables built-in KV-connector dependency libs into docker images
+# Flag enables build-in KV-connector dependency libs into docker images
 ARG INSTALL_KV_CONNECTORS=false
 
 #################### BASE BUILD IMAGE ####################
@@ -207,19 +207,6 @@ ARG SCCACHE_ENDPOINT
 ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
 ARG SCCACHE_REGION_NAME=us-west-2
 ARG SCCACHE_S3_NO_CREDENTIALS=0
-
-# Flag to control whether to use pre-built vLLM wheels
-ARG VLLM_USE_PRECOMPILED
-# TODO: in setup.py VLLM_USE_PRECOMPILED is sensitive to truthiness, it will take =0 as "true", this should be fixed
-ENV VLLM_USE_PRECOMPILED=""
-RUN if [ "${VLLM_USE_PRECOMPILED}" = "1" ]; then \
-        export VLLM_USE_PRECOMPILED=1 && \
-        echo "Using precompiled wheels"; \
-    else \
-        unset VLLM_USE_PRECOMPILED && \
-        echo "Leaving VLLM_USE_PRECOMPILED unset to build wheels from source"; \
-    fi
-
 # if USE_SCCACHE is set, use sccache to speed up compilation
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \

docker/Dockerfile.rocm_base

@@ -12,7 +12,7 @@ ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
 ARG FA_BRANCH="1a7f4dfa"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="916bf3c"
+ARG AITER_BRANCH="6487649"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base

docs/api/README.md

@@ -8,6 +8,7 @@ API documentation for vLLM's configuration classes.
 
 - [vllm.config.ModelConfig][]
 - [vllm.config.CacheConfig][]
+- [vllm.config.TokenizerPoolConfig][]
 - [vllm.config.LoadConfig][]
 - [vllm.config.ParallelConfig][]
 - [vllm.config.SchedulerConfig][]

docs/cli/README.md

@@ -1,7 +1,3 @@
----
-toc_depth: 4
----
-
 # vLLM CLI Guide
 
 The vllm command-line tool is used to run and manage vLLM models. You can start by viewing the help message with:
@@ -41,15 +37,8 @@ Start the vLLM OpenAI Compatible API server.
 
     # To search by keyword
     vllm serve --help=max
-
-    # To view full help with pager (less/more)
-    vllm serve --help=page
     ```
 
-### Options
-
---8<-- "docs/argparse/serve.md"
-
 ## chat
 
 Generate chat completions via the running API server.

docs/configuration/serve_args.md

@@ -5,7 +5,7 @@ The `vllm serve` command is used to launch the OpenAI-compatible server.
 ## CLI Arguments
 
 The `vllm serve` command is used to launch the OpenAI-compatible server.
-To see the available options, take a look at the [CLI Reference](../cli/README.md#options)!
+To see the available CLI arguments, run `vllm serve --help`!
 
 ## Configuration file
 

docs/contributing/model/basic.md

@@ -73,8 +73,6 @@ def forward(
     self,
     input_ids: torch.Tensor,
     positions: torch.Tensor,
-    intermediate_tensors: Optional[IntermediateTensors] = None,
-    inputs_embeds: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     ...
 ```

docs/deployment/frameworks/anyscale.md

@@ -3,15 +3,6 @@
 [](){ #deployment-anyscale }
 
 [Anyscale](https://www.anyscale.com) is a managed, multi-cloud platform developed by the creators of Ray.
-
-Anyscale automates the entire lifecycle of Ray clusters in your AWS, GCP, or Azure account, delivering the flexibility of open-source Ray
-without the operational overhead of maintaining Kubernetes control planes, configuring autoscalers, managing observability stacks, or manually managing head and worker nodes with helper scripts like <gh-file:examples/online_serving/run_cluster.sh>.
-
+It hosts Ray clusters inside your own AWS, GCP, or Azure account, delivering the flexibility of open-source Ray
+without the operational overhead of maintaining Kubernetes control planes, configuring autoscalers, or managing observability stacks.
 When serving large language models with vLLM, Anyscale can rapidly provision [production-ready HTTPS endpoints](https://docs.anyscale.com/examples/deploy-ray-serve-llms) or [fault-tolerant batch inference jobs](https://docs.anyscale.com/examples/ray-data-llm).
-
-## Production-ready vLLM on Anyscale quickstarts
-
-- [Offline batch inference](https://console.anyscale.com/template-preview/llm_batch_inference?utm_source=vllm_docs)
-- [Deploy vLLM services](https://console.anyscale.com/template-preview/llm_serving?utm_source=vllm_docs)
-- [Curate a dataset](https://console.anyscale.com/template-preview/audio-dataset-curation-llm-judge?utm_source=vllm_docs)
-- [Finetune an LLM](https://console.anyscale.com/template-preview/entity-recognition-with-llms?utm_source=vllm_docs)

docs/deployment/frameworks/open-webui.md

@@ -1,42 +1,26 @@
 # Open WebUI
 
-[Open WebUI](https://github.com/open-webui/open-webui) is an extensible, feature-rich,
-and user-friendly self-hosted AI platform designed to operate entirely offline.
-It supports various LLM runners like Ollama and OpenAI-compatible APIs,
-with built-in RAG capabilities, making it a powerful AI deployment solution.
+1. Install the [Docker](https://docs.docker.com/engine/install/)
 
-To get started with Open WebUI using vLLM, follow these steps:
+2. Start the vLLM server with the supported chat completion model, e.g.
 
-1. Install the [Docker](https://docs.docker.com/engine/install/).
+```bash
+vllm serve qwen/Qwen1.5-0.5B-Chat
+```
 
-2. Start the vLLM server with a supported chat completion model:
+1. Start the [Open WebUI](https://github.com/open-webui/open-webui) docker container (replace the vllm serve host and vllm serve port):
 
-    ```console
-    vllm serve Qwen/Qwen3-0.6B-Chat
-    ```
+```bash
+docker run -d -p 3000:8080 \
+--name open-webui \
+-v open-webui:/app/backend/data \
+-e OPENAI_API_BASE_URL=http://<vllm serve host>:<vllm serve port>/v1 \
+--restart always \
+ghcr.io/open-webui/open-webui:main
+```
 
-    !!! note
-        When starting the vLLM server, be sure to specify the host and port using the `--host` and `--port` flags.
-        For example:
+1. Open it in the browser: <http://open-webui-host:3000/>
 
-        ```console
-        python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000
-        ```
+On the top of the web page, you can see the model `qwen/Qwen1.5-0.5B-Chat`.
 
-3. Start the Open WebUI Docker container:
-
-    ```console
-    docker run -d \
-        --name open-webui \
-        -p 3000:8080 \
-        -v open-webui:/app/backend/data \
-        -e OPENAI_API_BASE_URL=http://0.0.0.0:8000/v1 \
-        --restart always \
-        ghcr.io/open-webui/open-webui:main
-    ```
-
-4. Open it in the browser: <http://open-webui-host:3000/>
-
-    At the top of the page, you should see the model `Qwen/Qwen3-0.6B-Chat`.
-
-    ![Web portal of model Qwen/Qwen3-0.6B-Chat](../../assets/deployment/open_webui.png)
+![](../../assets/deployment/open_webui.png)

docs/deployment/integrations/kuberay.md

@@ -1,20 +0,0 @@
-# KubeRay
-
-[KubeRay](https://github.com/ray-project/kuberay) provides a Kubernetes-native way to run vLLM workloads on Ray clusters.
-A Ray cluster can be declared in YAML, and the operator then handles pod scheduling, networking configuration, restarts, and blue-green deployments — all while preserving the familiar Kubernetes experience.
-
-## Why KubeRay instead of manual scripts?
-
-| Feature | Manual scripts | KubeRay |
-|---------|-----------------------------------------------------------|---------|
-| Cluster bootstrap | Manually SSH into every node and run a script | One command to create or update the whole cluster: `kubectl apply -f cluster.yaml` |
-| Autoscaling | Manual | Automatically patches CRDs for adjusting cluster size |
-| Upgrades | Tear down & re-create manually | Blue/green deployment updates supported |
-| Declarative config | Bash flags & environment variables | Git-ops-friendly YAML CRDs (RayCluster/RayService) |
-
-Using KubeRay reduces the operational burden and simplifies integration of Ray + vLLM with existing Kubernetes workflows (CI/CD, secrets, storage classes, etc.).
-
-## Learn more
-
-* ["Serve a Large Language Model using Ray Serve LLM on Kubernetes"](https://docs.ray.io/en/master/cluster/kubernetes/examples/rayserve-llm-example.html) - An end-to-end example of how to serve a model using vLLM, KubeRay, and Ray Serve.
-* [KubeRay documentation](https://docs.ray.io/en/latest/cluster/kubernetes/index.html)

docs/deployment/k8s.md

@@ -13,7 +13,6 @@ Alternatively, you can deploy vLLM to Kubernetes using any of the following:
 - [Helm](frameworks/helm.md)
 - [InftyAI/llmaz](integrations/llmaz.md)
 - [KServe](integrations/kserve.md)
-- [KubeRay](integrations/kuberay.md)
 - [kubernetes-sigs/lws](frameworks/lws.md)
 - [meta-llama/llama-stack](integrations/llamastack.md)
 - [substratusai/kubeai](integrations/kubeai.md)

docs/features/quantization/README.md

@@ -10,7 +10,6 @@ Contents:
 - [BitBLAS](bitblas.md)
 - [GGUF](gguf.md)
 - [GPTQModel](gptqmodel.md)
-- [INC](inc.md)
 - [INT4 W4A16](int4.md)
 - [INT8 W8A8](int8.md)
 - [FP8 W8A8](fp8.md)

docs/features/quantization/inc.md

@@ -1,56 +0,0 @@
----
-title: FP8 INC
----
-[](){ #inc }
-
-vLLM supports FP8 (8-bit floating point) weight and activation quantization using Intel® Neural Compressor (INC) on Intel® Gaudi® 2 and Intel® Gaudi® 3 AI accelerators.
-Currently, quantization is validated only in Llama models.
-
-Intel Gaudi supports quantization of various modules and functions, including, but not limited to `Linear`, `KVCache`, `Matmul` and `Softmax`. For more information, please refer to:
-[Supported Modules\\Supported Functions\\Custom Patched Modules](https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Quantization/Inference_Using_FP8.html#supported-modules).
-
-!!! note
-    Measurement files are required to run quantized models with vLLM on Gaudi accelerators. The FP8 model calibration procedure is described in the [vllm-hpu-extention](https://github.com/HabanaAI/vllm-hpu-extension/tree/main/calibration/README.md) package.
-
-!!! note
-    `QUANT_CONFIG` is an environment variable that points to the measurement or quantization [JSON config file](https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Quantization/Inference_Using_FP8.html#supported-json-config-file-options).
-    The measurement configuration file is used during the calibration procedure to collect measurements for a given model. The quantization configuration is used during inference.
-
-## Run Online Inference Using FP8
-
-Once you've completed the model calibration process and collected the measurements, you can run FP8 inference with vLLM using the following command:
-
-```bash
-export QUANT_CONFIG=/path/to/quant/config/inc/meta-llama-3.1-405b-instruct/maxabs_measure_g3.json
-vllm serve meta-llama/Llama-3.1-405B-Instruct --quantization inc --kv-cache-dtype fp8_inc --tensor_paralel_size 8
-```
-
-!!! tip
-    If you are just prototyping or testing your model with FP8, you can use the `VLLM_SKIP_WARMUP=true` environment variable to disable the warmup stage, which can take a long time. However, we do not recommend disabling this feature in production environments as it causes a significant performance drop.
-
-!!! tip
-    When using FP8 models, you may experience timeouts caused by the long compilation time of FP8 operations. To mitigate this problem, you can use the below environment variables:
-    `VLLM_ENGINE_ITERATION_TIMEOUT_S` - to adjust the vLLM server timeout. You can set the value in seconds, e.g., 600 equals 10 minutes.
-    `VLLM_RPC_TIMEOUT` - to adjust the RPC protocol timeout used by the OpenAI-compatible API. This value is in microseconds, e.g., 600000 equals 10 minutes.
-
-## Run Offline Inference Using FP8
-
-To run offline inference (after completing the model calibration process):
-
-* Set the "QUANT_CONFIG" environment variable to point to a JSON configuration file with QUANTIZE mode.
-* Pass `quantization=inc` and `kv_cache_dtype=fp8_inc` as parameters to the `LLM` object.
-* Call shutdown method of the model_executor at the end of the run.
-
-```python
-from vllm import LLM
-llm = LLM("llama3.1/Meta-Llama-3.1-8B-Instruct", quantization="inc", kv_cache_dtype="fp8_inc")
-...
-# Call llm.generate on the required prompts and sampling params.
-...
-llm.llm_engine.model_executor.shutdown()
-```
-
-## Device for the Model's Weights Uploading
-
-The unquantized weights are first loaded onto the CPU, then quantized and transferred to the target device (HPU) for model execution.
-This reduces the device memory footprint of model weights, as only quantized weights are stored in the device memory.

docs/features/quantization/supported_hardware.md

@@ -2,19 +2,18 @@
 
 The table below shows the compatibility of various quantization implementations with different hardware platforms in vLLM:
 
-| Implementation        | Volta   | Turing   | Ampere   | Ada   | Hopper   | AMD GPU   | Intel GPU   | Intel Gaudi | x86 CPU   | AWS Neuron   | Google TPU   |
-|-----------------------|---------|----------|----------|-------|----------|-----------|-------------|-------------|-----------|--------------|--------------|
-| AWQ                   | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌        | ✅︎          | ❌         | ✅︎        | ❌           | ❌           |
-| GPTQ                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎        | ❌        | ✅︎          | ❌         | ✅︎        | ❌           | ❌           |
-| Marlin (GPTQ/AWQ/FP8) | ❌      | ❌      | ✅︎       | ✅︎    | ✅︎       | ❌        | ❌          | ❌         | ❌        | ❌          | ❌           |
-| INT8 (W8A8)           | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌        | ❌          | ❌         | ✅︎        | ✅︎           | ✅︎            |
-| FP8 (W8A8)            | ❌      | ❌      | ❌       | ✅︎    | ✅︎      | ✅︎         | ❌          | ❌         | ❌        | ✅︎           | ❌           |
-| BitBLAS (GPTQ)        | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌         | ❌        | ❌          | ❌           |
-| AQLM                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌         | ❌        | ❌          | ❌           |
-| bitsandbytes          | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌         | ❌        | ❌          | ❌           |
-| DeepSpeedFP           | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌         | ❌        | ❌          | ❌           |
-| GGUF                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ✅︎         | ❌          | ❌         | ❌         | ❌          | ❌           |
-| INC (W8A8)            | ❌      | ❌      | ❌      | ❌    | ❌      | ❌        | ❌          | ✅︎         | ❌         | ❌           | ❌          |
+| Implementation        | Volta   | Turing   | Ampere   | Ada   | Hopper   | AMD GPU   | Intel GPU   | x86 CPU   | AWS Neuron   | Google TPU   |
+|-----------------------|---------|----------|----------|-------|----------|-----------|-------------|-----------|------------------|--------------|
+| AWQ                   | ❌       | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        | ❌                | ❌            |
+| GPTQ                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        | ❌                | ❌            |
+| Marlin (GPTQ/AWQ/FP8) | ❌       | ❌        | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ❌         | ❌                | ❌            |
+| INT8 (W8A8)           | ❌       | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ✅︎        | ✅︎                | ✅︎           |
+| FP8 (W8A8)            | ❌       | ❌        | ❌        | ✅︎    | ✅︎       | ✅︎        | ❌           | ❌         | ✅︎                | ❌            |
+| BitBLAS (GPTQ)        | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ❌         | ❌                | ❌            |
+| AQLM                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ❌         | ❌                | ❌            |
+| bitsandbytes          | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ❌         | ❌                | ❌            |
+| DeepSpeedFP           | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌           | ❌         | ❌                | ❌            |
+| GGUF                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ✅︎        | ❌           | ❌         | ❌                | ❌            |
 
 - Volta refers to SM 7.0, Turing to SM 7.5, Ampere to SM 8.0/8.6, Ada to SM 8.9, and Hopper to SM 9.0.
 - ✅︎ indicates that the quantization method is supported on the specified hardware.

docs/features/tool_calling.md

@@ -103,7 +103,9 @@ When tool_choice='required' is set, the model is guaranteed to generate one or m
 
 vLLM supports the `tool_choice='none'` option in the chat completion API. When this option is set, the model will not generate any tool calls and will respond with regular text content only, even if tools are defined in the request.
 
-However, when `tool_choice='none'` is specified, vLLM includes tool definitions from the prompt.
+By default, when `tool_choice='none'` is specified, vLLM excludes tool definitions from the prompt to optimize context usage. To include tool definitions even with `tool_choice='none'`, use the `--expand-tools-even-if-tool-choice-none` option.
+
+Note: This behavior will change in v0.10.0, where tool definitions will be included by default even with `tool_choice='none'`.
 
 ## Automatic Function Calling
 
@@ -280,14 +282,6 @@ Supported models:
 
 Flags: `--tool-call-parser deepseek_v3 --chat-template {see_above}`
 
-### Kimi-K2 Models (`kimi_k2`)
-
-Supported models:
-
-* `moonshotai/Kimi-K2-Instruct`
-
-Flags: `--tool-call-parser kimi_k2`
-
 ### Models with Pythonic Tool Calls (`pythonic`)
 
 A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.

docs/getting_started/installation/intel_gaudi.md

@@ -28,7 +28,7 @@ To verify that the Intel Gaudi software was correctly installed, run:
 hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
 apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core, habanalabs-thunk and habanalabs-container-runtime are installed
 pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
-pip list | grep neural # verify that neural_compressor_pt is installed
+pip list | grep neural # verify that neural_compressor is installed
 ```
 
 Refer to [Intel Gaudi Software Stack Verification](https://docs.habana.ai/en/latest/Installation_Guide/SW_Verification.html#platform-upgrade)
@@ -120,13 +120,12 @@ docker run \
 - Inference with [HPU Graphs](https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html)
   for accelerating low-batch latency and throughput
 - Attention with Linear Biases (ALiBi)
-- INC quantization
 
 ### Unsupported features
 
 - Beam search
 - LoRA adapters
-- AWQ quantization
+- Quantization
 - Prefill chunking (mixed-batch inferencing)
 
 ### Supported configurations
@@ -134,20 +133,36 @@ docker run \
 The following configurations have been validated to function with
 Gaudi2 devices. Configurations that are not listed may or may not work.
 
-| Model | TP Size| dtype | Sampling |
-|-------|--------|--------|----------|
-| [meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3.1-8B](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct) | 1, 2, 8 | BF16 | Random / Greedy |
-| [meta-llama/Llama-2-70b](https://huggingface.co/meta-llama/Llama-2-70b) | 8 | BF16 | Random / Greedy |
-| [meta-llama/Llama-2-70b-chat-hf](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) | 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B) | 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3.1-70B](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B) | 8 | BF16 | Random / Greedy |
-| [meta-llama/Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct) | 8 | BF16 | Random / Greedy |
+- [meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3.1-8B](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct)
+  on single HPU, or with tensor parallelism on 2x and 8x HPU, BF16
+  datatype with random or greedy sampling
+- [meta-llama/Llama-2-70b](https://huggingface.co/meta-llama/Llama-2-70b)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
+- [meta-llama/Llama-2-70b-chat-hf](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3.1-70B](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
+- [meta-llama/Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct)
+  with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
 
 ## Performance tuning
 

docs/mkdocs/hooks/generate_argparse.py

@@ -16,7 +16,6 @@ sys.modules["blake3"] = MagicMock()
 sys.modules["vllm._C"] = MagicMock()
 
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs  # noqa: E402
-from vllm.entrypoints.openai.cli_args import make_arg_parser  # noqa: E402
 from vllm.utils import FlexibleArgumentParser  # noqa: E402
 
 logger = logging.getLogger("mkdocs")
@@ -25,18 +24,15 @@ logger = logging.getLogger("mkdocs")
 class MarkdownFormatter(HelpFormatter):
     """Custom formatter that generates markdown for argument groups."""
 
-    def __init__(self, prog, starting_heading_level=3):
+    def __init__(self, prog):
         super().__init__(prog,
                          max_help_position=float('inf'),
                          width=float('inf'))
-        self._section_heading_prefix = "#" * starting_heading_level
-        self._argument_heading_prefix = "#" * (starting_heading_level + 1)
         self._markdown_output = []
 
     def start_section(self, heading):
         if heading not in {"positional arguments", "options"}:
-            heading_md = f"\n{self._section_heading_prefix} {heading}\n\n"
-            self._markdown_output.append(heading_md)
+            self._markdown_output.append(f"\n### {heading}\n\n")
 
     def end_section(self):
         pass
@@ -50,13 +46,9 @@ class MarkdownFormatter(HelpFormatter):
 
     def add_arguments(self, actions):
         for action in actions:
-            if (len(action.option_strings) == 0
-                    or "--help" in action.option_strings):
-                continue
 
             option_strings = f'`{"`, `".join(action.option_strings)}`'
-            heading_md = f"{self._argument_heading_prefix} {option_strings}\n\n"
-            self._markdown_output.append(heading_md)
+            self._markdown_output.append(f"#### {option_strings}\n\n")
 
             if choices := action.choices:
                 choices = f'`{"`, `".join(str(c) for c in choices)}`'
@@ -89,14 +81,6 @@ def create_parser(cls, **kwargs) -> FlexibleArgumentParser:
         return cls.add_cli_args(parser, **kwargs)
 
 
-def create_serve_parser() -> FlexibleArgumentParser:
-    """Create a parser for the serve command with markdown formatting."""
-    parser = FlexibleArgumentParser()
-    parser.formatter_class = lambda prog: MarkdownFormatter(
-        prog, starting_heading_level=4)
-    return make_arg_parser(parser)
-
-
 def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
     logger.info("Generating argparse documentation")
     logger.debug("Root directory: %s", ROOT_DIR.resolve())
@@ -111,7 +95,6 @@ def on_startup(command: Literal["build", "gh-deploy", "serve"], dirty: bool):
         "engine_args": create_parser(EngineArgs),
         "async_engine_args": create_parser(AsyncEngineArgs,
                                            async_args_only=True),
-        "serve": create_serve_parser(),
     }
 
     # Generate documentation for each parser

docs/models/supported_models.md

@@ -316,7 +316,6 @@ Specified using `--task generate`.
 | `AquilaForCausalLM` | Aquila, Aquila2 | `BAAI/Aquila-7B`, `BAAI/AquilaChat-7B`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `ArcticForCausalLM` | Arctic | `Snowflake/snowflake-arctic-base`, `Snowflake/snowflake-arctic-instruct`, etc. | | ✅︎ | ✅︎ |
 | `BaiChuanForCausalLM` | Baichuan2, Baichuan | `baichuan-inc/Baichuan2-13B-Chat`, `baichuan-inc/Baichuan-7B`, etc. | ✅︎ | ✅︎ | ✅︎ |
-| `BailingMoeForCausalLM` | Ling | `inclusionAI/Ling-lite-1.5`, `inclusionAI/Ling-plus`, etc. | | ✅︎ | ✅︎ |
 | `BambaForCausalLM` | Bamba | `ibm-ai-platform/Bamba-9B-fp8`, `ibm-ai-platform/Bamba-9B` | ✅︎ | ✅︎ | ✅︎ |
 | `BloomForCausalLM` | BLOOM, BLOOMZ, BLOOMChat | `bigscience/bloom`, `bigscience/bloomz`, etc. | | ✅︎ | |
 | `BartForConditionalGeneration` | BART | `facebook/bart-base`, `facebook/bart-large-cnn`, etc. | | | |
@@ -375,7 +374,6 @@ Specified using `--task generate`.
 | `Phi3ForCausalLM` | Phi-4, Phi-3 | `microsoft/Phi-4-mini-instruct`, `microsoft/Phi-4`, `microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, `microsoft/Phi-3-medium-128k-instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Phi3SmallForCausalLM` | Phi-3-Small | `microsoft/Phi-3-small-8k-instruct`, `microsoft/Phi-3-small-128k-instruct`, etc. | | ✅︎ | ✅︎ |
 | `PhiMoEForCausalLM` | Phi-3.5-MoE | `microsoft/Phi-3.5-MoE-instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
-| `Phi4FlashForCausalLM` | Phi-4-mini-flash-reasoning | `microsoft/microsoft/Phi-4-mini-instruct`, etc. | | | |
 | `PersimmonForCausalLM` | Persimmon | `adept/persimmon-8b-base`, `adept/persimmon-8b-chat`, etc. | | ✅︎ | ✅︎ |
 | `Plamo2ForCausalLM` | PLaMo2 | `pfnet/plamo-2-1b`, `pfnet/plamo-2-8b`, etc. | | | |
 | `QWenLMHeadModel` | Qwen | `Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc. | ✅︎ | ✅︎ | ✅︎ |
@@ -581,14 +579,14 @@ Specified using `--task generate`.
 | `KeyeForConditionalGeneration` | Keye-VL-8B-Preview | T + I<sup>E+</sup> + V<sup>E+</sup> | `Kwai-Keye/Keye-VL-8B-Preview` | | | ✅︎ |
 | `KimiVLForConditionalGeneration` | Kimi-VL-A3B-Instruct, Kimi-VL-A3B-Thinking | T + I<sup>+</sup> | `moonshotai/Kimi-VL-A3B-Instruct`, `moonshotai/Kimi-VL-A3B-Thinking` | | | ✅︎ |
 | `Llama4ForConditionalGeneration` | Llama 4 | T + I<sup>+</sup> | `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc. | | ✅︎ | ✅︎ |
-| `LlavaForConditionalGeneration` | LLaVA-1.5, Pixtral (HF Transformers) | T + I<sup>E+</sup> | `llava-hf/llava-1.5-7b-hf`, `TIGER-Lab/Mantis-8B-siglip-llama3` (see note), `mistral-community/pixtral-12b`, etc. | | ✅︎ | ✅︎ |
+| `LlavaForConditionalGeneration` | LLaVA-1.5 | T + I<sup>E+</sup> | `llava-hf/llava-1.5-7b-hf`, `TIGER-Lab/Mantis-8B-siglip-llama3` (see note), etc. | | ✅︎ | ✅︎ |
 | `LlavaNextForConditionalGeneration` | LLaVA-NeXT | T + I<sup>E+</sup> | `llava-hf/llava-v1.6-mistral-7b-hf`, `llava-hf/llava-v1.6-vicuna-7b-hf`, etc. | | ✅︎ | ✅︎ |
 | `LlavaNextVideoForConditionalGeneration` | LLaVA-NeXT-Video | T + V | `llava-hf/LLaVA-NeXT-Video-7B-hf`, etc. | | ✅︎ | ✅︎ |
 | `LlavaOnevisionForConditionalGeneration` | LLaVA-Onevision | T + I<sup>+</sup> + V<sup>+</sup> | `llava-hf/llava-onevision-qwen2-7b-ov-hf`, `llava-hf/llava-onevision-qwen2-0.5b-ov-hf`, etc. | | ✅︎ | ✅︎ |
 | `MiniCPMO` | MiniCPM-O | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>E+</sup> | `openbmb/MiniCPM-o-2_6`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `MiniCPMV` | MiniCPM-V | T + I<sup>E+</sup> + V<sup>E+</sup> | `openbmb/MiniCPM-V-2` (see note), `openbmb/MiniCPM-Llama3-V-2_5`, `openbmb/MiniCPM-V-2_6`, etc. | ✅︎ | | ✅︎ |
 | `MiniMaxVL01ForConditionalGeneration` | MiniMax-VL | T + I<sup>E+</sup> | `MiniMaxAI/MiniMax-VL-01`, etc. | | ✅︎ | ✅︎ |
-| `Mistral3ForConditionalGeneration` | Mistral3 (HF Transformers) | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `Mistral3ForConditionalGeneration` | Mistral3 | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `MllamaForConditionalGeneration` | Llama 3.2 | T + I<sup>+</sup> | `meta-llama/Llama-3.2-90B-Vision-Instruct`, `meta-llama/Llama-3.2-11B-Vision`, etc. | | | |
 | `MolmoForCausalLM` | Molmo | T + I<sup>+</sup> | `allenai/Molmo-7B-D-0924`, `allenai/Molmo-7B-O-0924`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `NVLM_D_Model` | NVLM-D 1.0 | T + I<sup>+</sup> | `nvidia/NVLM-D-72B`, etc. | | ✅︎ | ✅︎ |
@@ -596,7 +594,7 @@ Specified using `--task generate`.
 | `PaliGemmaForConditionalGeneration` | PaliGemma, PaliGemma 2 | T + I<sup>E</sup> | `google/paligemma-3b-pt-224`, `google/paligemma-3b-mix-224`, `google/paligemma2-3b-ft-docci-448`, etc. | | ✅︎ | ⚠️ |
 | `Phi3VForCausalLM` | Phi-3-Vision, Phi-3.5-Vision | T + I<sup>E+</sup> | `microsoft/Phi-3-vision-128k-instruct`, `microsoft/Phi-3.5-vision-instruct`, etc. | | ✅︎ | ✅︎ |
 | `Phi4MMForCausalLM` | Phi-4-multimodal | T + I<sup>+</sup> / T + A<sup>+</sup> / I<sup>+</sup> + A<sup>+</sup> | `microsoft/Phi-4-multimodal-instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
-| `PixtralForConditionalGeneration` | Mistral 3 (Mistral format), Pixtral (Mistral format) | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, `mistralai/Pixtral-12B-2409`, etc. | | ✅︎ | ✅︎ |
+| `PixtralForConditionalGeneration` | Pixtral | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, `mistral-community/pixtral-12b`, etc. | | ✅︎ | ✅︎ |
 | `QwenVLForConditionalGeneration`<sup>^</sup> | Qwen-VL | T + I<sup>E+</sup> | `Qwen/Qwen-VL`, `Qwen/Qwen-VL-Chat`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen2AudioForConditionalGeneration` | Qwen2-Audio | T + A<sup>+</sup> | `Qwen/Qwen2-Audio-7B-Instruct` | | ✅︎ | ✅︎ |
 | `Qwen2VLForConditionalGeneration` | QVQ, Qwen2-VL | T + I<sup>E+</sup> + V<sup>E+</sup> | `Qwen/QVQ-72B-Preview`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |

docs/serving/data_parallel_deployment.md

@@ -57,20 +57,12 @@ vllm serve $MODEL --headless --data-parallel-size 4 --data-parallel-size-local 4
                   --data-parallel-address 10.99.48.128 --data-parallel-rpc-port 13345
 ```
 
-This DP mode can also be used with Ray by specifying `--data-parallel-backend=ray`:
+This DP mode can also be used with Ray, in which case only a single launch command is needed irrespective of the number of nodes:
 
 ```bash
-vllm serve $MODEL --data-parallel-size 4 --data-parallel-size-local 2 \
-                  --data-parallel-backend=ray
+vllm serve $MODEL --data-parallel-size 16 --tensor-parallel-size 2 --data-parallel-backend=ray
 ```
 
-There are several notable differences when using Ray:
-
-- A single launch command (on any node) is needed to start all local and remote DP ranks, therefore it is more convenient compared to launching on each node
-- There is no need to specify `--data-parallel-address`, and the node where the command is run is used as `--data-parallel-address`
-- There is no need to specify `--data-parallel-rpc-port`
-- Remote DP ranks will be allocated based on node resources of the Ray cluster
-
 Currently, the internal DP load balancing is done within the API server process(es) and is based on the running and waiting queues in each of the engines. This could be made more sophisticated in future by incorporating KV cache aware logic.
 
 When deploying large DP sizes using this method, the API server process can become a bottleneck. In this case, the orthogonal `--api-server-count` command line option can be used to scale this out (for example `--api-server-count=4`). This is transparent to users - a single HTTP endpoint / port is still exposed. Note that this API server scale-out is "internal" and still confined to the "head" node.

docs/usage/v1_guide.md

@@ -106,13 +106,14 @@ to enable simultaneous generation and embedding using the same engine instance i
 
 Models using selective state-space mechanisms instead of standard transformer attention are partially supported.
 Models that use Mamba-2 layers (e.g., `Mamba2ForCausalLM`) are supported, but models that use older Mamba-1 layers
-(e.g., `MambaForCausalLM`, `JambaForCausalLM`) are not yet supported. Please note that these models currently require
+(e.g., `MambaForCausalLM`, `JambaForCausalLM`) are not yet suported. Please note that these models currently require
 enforcing eager mode and disabling prefix caching in V1.
 
 Models that combine Mamba-2 layers with standard attention layers are also supported (e.g., `BambaForCausalLM`,
 `Zamba2ForCausalLM`, `NemotronHForCausalLM`, `FalconH1ForCausalLM` and `GraniteMoeHybridForCausalLM`). Please note that
 these models currently require enforcing eager mode, disabling prefix caching, and using the FlashInfer attention
-backend in V1.
+backend in V1. It is also necessary to pass a non-standard block size for attention layers (this is not possible
+using the `vllm serve` CLI yet).
 
 #### Encoder-Decoder Models
 

examples/offline_inference/audio_language.py

@@ -10,7 +10,7 @@ on HuggingFace model repository.
 
 import os
 from dataclasses import asdict
-from typing import Any, NamedTuple, Optional
+from typing import NamedTuple, Optional
 
 from huggingface_hub import snapshot_download
 from transformers import AutoTokenizer
@@ -30,9 +30,7 @@ question_per_audio_count = {
 
 class ModelRequestData(NamedTuple):
     engine_args: EngineArgs
-    prompt: Optional[str] = None
-    prompt_token_ids: Optional[dict[str, list[int]]] = None
-    multi_modal_data: Optional[dict[str, Any]] = None
+    prompt: str
     stop_token_ids: Optional[list[int]] = None
     lora_requests: Optional[list[LoRARequest]] = None
 
@@ -42,60 +40,6 @@ class ModelRequestData(NamedTuple):
 # Unless specified, these settings have been tested to work on a single L4.
 
 
-# Voxtral
-def run_voxtral(question: str, audio_count: int) -> ModelRequestData:
-    from mistral_common.audio import Audio
-    from mistral_common.protocol.instruct.messages import (
-        AudioChunk,
-        RawAudio,
-        TextChunk,
-        UserMessage,
-    )
-    from mistral_common.protocol.instruct.request import ChatCompletionRequest
-    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
-
-    model_name = "mistralai/Voxtral-Mini-3B-2507"
-    tokenizer = MistralTokenizer.from_hf_hub(model_name)
-
-    engine_args = EngineArgs(
-        model=model_name,
-        max_model_len=8192,
-        max_num_seqs=2,
-        limit_mm_per_prompt={"audio": audio_count},
-        config_format="mistral",
-        load_format="mistral",
-        tokenizer_mode="mistral",
-        enforce_eager=True,
-        enable_chunked_prefill=False,
-    )
-
-    text_chunk = TextChunk(text=question)
-    audios = [
-        Audio.from_file(str(audio_assets[i].get_local_path()), strict=False)
-        for i in range(audio_count)
-    ]
-    audio_chunks = [
-        AudioChunk(input_audio=RawAudio.from_audio(audio)) for audio in audios
-    ]
-
-    messages = [UserMessage(content=[*audio_chunks, text_chunk])]
-
-    req = ChatCompletionRequest(messages=messages, model=model_name)
-
-    tokens = tokenizer.encode_chat_completion(req)
-    prompt_ids, audios = tokens.tokens, tokens.audios
-
-    audios_and_sr = [(au.audio_array, au.sampling_rate) for au in audios]
-
-    multi_modal_data = {"audio": audios_and_sr}
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        prompt_token_ids=prompt_ids,
-        multi_modal_data=multi_modal_data,
-    )
-
-
 # Granite Speech
 def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
     # NOTE - the setting in this example are somehat different than what is
@@ -299,7 +243,6 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
 
 
 model_example_map = {
-    "voxtral": run_voxtral,
     "granite_speech": run_granite_speech,
     "minicpmo": run_minicpmo,
     "phi4_mm": run_phi4mm,
@@ -368,24 +311,16 @@ def main(args):
         temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
     )
 
-    mm_data = req_data.multi_modal_data
-    if not mm_data:
-        mm_data = {}
-        if audio_count > 0:
-            mm_data = {
-                "audio": [
-                    asset.audio_and_sample_rate for asset in audio_assets[:audio_count]
-                ]
-            }
+    mm_data = {}
+    if audio_count > 0:
+        mm_data = {
+            "audio": [
+                asset.audio_and_sample_rate for asset in audio_assets[:audio_count]
+            ]
+        }
 
     assert args.num_prompts > 0
-    inputs = {"multi_modal_data": mm_data}
-
-    if req_data.prompt:
-        inputs["prompt"] = req_data.prompt
-    else:
-        inputs["prompt_token_ids"] = req_data.prompt_token_ids
-
+    inputs = {"prompt": req_data.prompt, "multi_modal_data": mm_data}
     if args.num_prompts > 1:
         # Batch inference
         inputs = [inputs] * args.num_prompts

examples/offline_inference/rlhf.py

@@ -1,31 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
-Demonstrates reinforcement learning from human feedback (RLHF) using vLLM and Ray.
-
-The script separates training and inference workloads onto distinct GPUs
-so that Ray can manage process placement and inter-process communication.
-A Hugging Face Transformer model occupies GPU 0 for training, whereas a
-tensor-parallel vLLM inference engine occupies GPU 1–2.
-
-The example performs the following steps:
-
-* Load the training model on GPU 0.
-* Split the inference model across GPUs 1–2 using vLLM's tensor parallelism
-  and Ray placement groups.
-* Generate text from a list of prompts using the inference engine.
-* Update the weights of the training model and broadcast the updated weights
-  to the inference engine by using a Ray collective RPC group. Note that
-  for demonstration purposes we simply zero out the weights.
-
-For a production-ready implementation that supports multiple training and
-inference replicas, see the OpenRLHF framework:
-https://github.com/OpenRLHF/OpenRLHF
-
-This example assumes a single-node cluster with three GPUs, but Ray
-supports multi-node clusters. vLLM expects the GPUs are only used for vLLM
-workloads. Residual GPU activity interferes with vLLM memory profiling and
-causes unexpected behavior.
+a simple demonstration of RLHF with vLLM, inspired by
+the OpenRLHF framework https://github.com/OpenRLHF/OpenRLHF .
+It follows the design that, training processes and inference processes
+are different, and they live on different GPUs.
+Training processes send prompts to inference processes to generate data,
+and also synchronize the weights of the model by broadcasting the weights
+from the training process to the inference process.
+Note that this is a simple demonstration of one training instance and one
+inference instance. In practice, there could be multiple training instances
+and multiple inference instances. For the full implementation, please refer
+to the OpenRLHF framework.
 """
 
 import os
@@ -42,27 +28,29 @@ from vllm.utils import get_ip, get_open_port
 
 
 class MyLLM(LLM):
-    """Configure the vLLM worker for Ray placement group execution."""
-
     def __init__(self, *args, **kwargs):
-        # Remove the top-level CUDA_VISIBLE_DEVICES variable set by Ray
-        # so that vLLM can manage its own device placement within the worker.
+        # a hack to make the script work.
+        # stop ray from manipulating CUDA_VISIBLE_DEVICES
+        # at the top-level
         os.environ.pop("CUDA_VISIBLE_DEVICES", None)
         super().__init__(*args, **kwargs)
 
 
-# Load the OPT-125M model onto GPU 0 for the training workload.
+"""
+Start the training process, here we use huggingface transformers 
+as an example to hold a model on GPU 0.
+"""
+
 train_model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m")
 train_model.to("cuda:0")
-
-# Initialize Ray and set the visible devices. The vLLM engine will
-# be placed on GPUs 1 and 2.
+"""
+Start the inference process, here we use vLLM to hold a model on GPU 1 and 
+GPU 2. For the details on how to use ray, please refer to the ray 
+documentation https://docs.ray.io/en/latest/ .
+"""
 os.environ["CUDA_VISIBLE_DEVICES"] = "1,2"
 ray.init()
 
-# Create a placement group that reserves GPU 1–2 for the vLLM inference engine.
-# Learn more about Ray placement groups:
-# https://docs.ray.io/en/latest/placement-groups.html
 pg_inference = placement_group([{"GPU": 1, "CPU": 0}] * 2)
 ray.get(pg_inference.ready())
 scheduling_inference = PlacementGroupSchedulingStrategy(
@@ -70,9 +58,10 @@ scheduling_inference = PlacementGroupSchedulingStrategy(
     placement_group_capture_child_tasks=True,
     placement_group_bundle_index=0,
 )
-
-# Launch the vLLM inference engine. The `enforce_eager` flag reduces
-# start-up latency.
+"""
+launch the vLLM inference engine.
+here we use `enforce_eager` to reduce the start time.
+"""
 llm = ray.remote(
     num_cpus=0,
     num_gpus=0,
@@ -85,7 +74,7 @@ llm = ray.remote(
     distributed_executor_backend="ray",
 )
 
-# Generate text from the prompts.
+# Generate texts from the prompts.
 prompts = [
     "Hello, my name is",
     "The president of the United States is",
@@ -104,8 +93,8 @@ for output in outputs:
     print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
     print("-" * 50)
 
-# Set up the communication channel between the training process and the
-# inference engine.
+# set up the communication between the training process
+# and the inference engine.
 master_address = get_ip()
 master_port = get_open_port()
 
@@ -118,23 +107,21 @@ model_update_group = stateless_init_process_group(
 )
 ray.get(handle)
 
-# Simulate a training step by zeroing out all model weights.
-# In a real RLHF training loop the weights would be updated using the gradient
-# from an RL objective such as PPO on a reward model.
+# simulate training, modify the weights of the model.
 for name, p in train_model.named_parameters():
     p.data.zero_()
 
-# Synchronize the updated weights to the inference engine.
+# sync weight from the training process to the inference engine.
 for name, p in train_model.named_parameters():
     handle = llm.collective_rpc.remote("update_weight", args=(name, p.dtype, p.shape))
     model_update_group.broadcast(p, src=0, stream=torch.cuda.current_stream())
     ray.get(handle)
 
-# Verify that the inference weights have been updated.
+# check if the weights are updated.
 assert all(ray.get(llm.collective_rpc.remote("check_weights_changed")))
 
-# Generate text with the updated model. The output is expected to be nonsense
-# because the weights are zero.
+# use the updated model to generate texts, they will be nonsense
+# because the weights are all zeros.
 outputs_updated = ray.get(llm.generate.remote(prompts, sampling_params))
 print("-" * 50)
 for output in outputs_updated:

examples/offline_inference/spec_decode.py

@@ -84,7 +84,6 @@ def main():
         gpu_memory_utilization=0.8,
         speculative_config=speculative_config,
         disable_log_stats=False,
-        max_model_len=16384,
     )
 
     sampling_params = SamplingParams(temperature=args.temp, max_tokens=args.output_len)

examples/online_serving/run_cluster.sh

@@ -1,81 +1,35 @@
 #!/bin/bash
-#
-# Launch a Ray cluster inside Docker for vLLM inference.
-#
-# This script can start either a head node or a worker node, depending on the
-# --head or --worker flag provided as the third positional argument.
-#
-# Usage:
-# 1. Designate one machine as the head node and execute:
-#    bash run_cluster.sh \
-#         vllm/vllm-openai \
-#         <head_node_ip> \
-#         --head \
-#         /abs/path/to/huggingface/cache \
-#         -e VLLM_HOST_IP=<head_node_ip>
-#
-# 2. On every worker machine, execute:
-#    bash run_cluster.sh \
-#         vllm/vllm-openai \
-#         <head_node_ip> \
-#         --worker \
-#         /abs/path/to/huggingface/cache \
-#         -e VLLM_HOST_IP=<worker_node_ip>
-# 
-# Each worker requires a unique VLLM_HOST_IP value.
-# Keep each terminal session open. Closing a session stops the associated Ray
-# node and thereby shuts down the entire cluster.
-# Every machine must be reachable at the supplied IP address.
-#
-# The container is named "node-<random_suffix>". To open a shell inside
-# a container after launch, use:
-#       docker exec -it node-<random_suffix> /bin/bash
-#
-# Then, you can execute vLLM commands on the Ray cluster as if it were a
-# single machine, e.g. vllm serve ...
-#
-# To stop the container, use:
-#       docker stop node-<random_suffix>
 
-# Check for minimum number of required arguments.
+# Check for minimum number of required arguments
 if [ $# -lt 4 ]; then
-    echo "Usage: $0 docker_image head_node_ip --head|--worker path_to_hf_home [additional_args...]"
+    echo "Usage: $0 docker_image head_node_address --head|--worker path_to_hf_home [additional_args...]"
     exit 1
 fi
 
-# Extract the mandatory positional arguments and remove them from $@.
+# Assign the first three arguments and shift them away
 DOCKER_IMAGE="$1"
 HEAD_NODE_ADDRESS="$2"
-NODE_TYPE="$3"  # Should be --head or --worker.
+NODE_TYPE="$3"  # Should be --head or --worker
 PATH_TO_HF_HOME="$4"
 shift 4
 
-# Preserve any extra arguments so they can be forwarded to Docker.
+# Additional arguments are passed directly to the Docker command
 ADDITIONAL_ARGS=("$@")
 
-# Validate the NODE_TYPE argument.
+# Validate node type
 if [ "${NODE_TYPE}" != "--head" ] && [ "${NODE_TYPE}" != "--worker" ]; then
     echo "Error: Node type must be --head or --worker"
     exit 1
 fi
 
-# Generate a unique container name with random suffix.
-# Docker container names must be unique on each host.
-# The random suffix allows multiple Ray containers to run simultaneously on the same machine,
-# for example, on a multi-GPU machine.
-CONTAINER_NAME="node-${RANDOM}"
-
-# Define a cleanup routine that removes the container when the script exits.
-# This prevents orphaned containers from accumulating if the script is interrupted.
+# Define a function to cleanup on EXIT signal
 cleanup() {
-    docker stop "${CONTAINER_NAME}"
-    docker rm "${CONTAINER_NAME}"
+    docker stop node
+    docker rm node
 }
 trap cleanup EXIT
 
-# Build the Ray start command based on the node role.
-# The head node manages the cluster and accepts connections on port 6379, 
-# while workers connect to the head's address.
+# Command setup for head or worker node
 RAY_START_CMD="ray start --block"
 if [ "${NODE_TYPE}" == "--head" ]; then
     RAY_START_CMD+=" --head --port=6379"
@@ -83,15 +37,11 @@ else
     RAY_START_CMD+=" --address=${HEAD_NODE_ADDRESS}:6379"
 fi
 
-# Launch the container with the assembled parameters.
-# --network host: Allows Ray nodes to communicate directly via host networking
-# --shm-size 10.24g: Increases shared memory
-# --gpus all: Gives container access to all GPUs on the host
-# -v HF_HOME: Mounts HuggingFace cache to avoid re-downloading models
+# Run the docker command with the user specified parameters and additional arguments
 docker run \
     --entrypoint /bin/bash \
     --network host \
-    --name "${CONTAINER_NAME}" \
+    --name node \
     --shm-size 10.24g \
     --gpus all \
     -v "${PATH_TO_HF_HOME}:/root/.cache/huggingface" \

mkdocs.yaml

@@ -61,7 +61,6 @@ plugins:
   - search
   - autorefs
   - awesome-nav
-  - glightbox
   # For API reference generation
   - api-autonav:
       modules: ["vllm"]

pyproject.toml

@@ -174,186 +174,3 @@ respect-ignore-files = true
 
 [tool.ty.environment]
 python = "./.venv"
-
-[tool.typos.files]
-# these files may be written in non english words
-extend-exclude = ["tests/models/fixtures/*", "tests/prompts/*",
-    "benchmarks/sonnet.txt", "tests/lora/data/*", "build/*",
-    "vllm/third_party/*"]
-ignore-hidden = true
-ignore-files = true
-ignore-dot = true
-ignore-vcs = true
-ignore-global = true
-ignore-parent = true
-
-[tool.typos.default]
-binary = false
-check-filename = false
-check-file = true
-unicode = true
-ignore-hex = true
-identifier-leading-digits = false
-locale = "en"
-extend-ignore-identifiers-re = ["NVML_*", ".*Unc.*", ".*_thw",
-    ".*UE8M0.*", ".*[UE4M3|ue4m3].*", ".*eles.*",
-     ".*[Tt]h[rR].*"]
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.default.extend-identifiers]
-bbc5b7ede = "bbc5b7ede"
-womens_doubles = "womens_doubles"
-v_2nd = "v_2nd"
-# splitted_input = "splitted_input"
-NOOPs = "NOOPs"
-typ = "typ"
-nin_shortcut = "nin_shortcut"
-UperNetDecoder = "UperNetDecoder"
-subtile = "subtile"
-cudaDevAttrMaxSharedMemoryPerBlockOptin = "cudaDevAttrMaxSharedMemoryPerBlockOptin"
-SFOuput = "SFOuput"
-# huggingface transformers repo uses these words
-depthwise_seperable_out_channel = "depthwise_seperable_out_channel"
-DepthWiseSeperableConv1d = "DepthWiseSeperableConv1d"
-depthwise_seperable_CNN = "depthwise_seperable_CNN"
-
-[tool.typos.default.extend-words]
-iy = "iy"
-tendencias = "tendencias"
-# intel cpu features
-tme = "tme"
-dout = "dout"
-Pn = "Pn"
-arange = "arange"
-
-[tool.typos.type.py]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.py.extend-identifiers]
-arange = "arange"
-NDArray = "NDArray"
-EOFError = "EOFError"
-fo = "fo"
-ba = "ba"
-
-[tool.typos.type.py.extend-words]
-
-[tool.typos.type.cpp]
-extend-glob = ["*.cu"]
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.cpp.extend-identifiers]
-countr_one = "countr_one"
-k_ot = "k_ot"
-ot = "ot"
-
-[tool.typos.type.cpp.extend-words]
-
-[tool.typos.type.rust]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.rust.extend-identifiers]
-flate2 = "flate2"
-
-[tool.typos.type.rust.extend-words]
-ser = "ser"
-
-[tool.typos.type.lock]
-extend-glob = []
-check-file = false
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.lock.extend-identifiers]
-
-[tool.typos.type.lock.extend-words]
-
-[tool.typos.type.jl]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.jl.extend-identifiers]
-
-[tool.typos.type.jl.extend-words]
-modul = "modul"
-egals = "egals"
-usig = "usig"
-egal = "egal"
-
-[tool.typos.type.go]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.go.extend-identifiers]
-flate = "flate"
-
-[tool.typos.type.go.extend-words]
-
-[tool.typos.type.css]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.css.extend-identifiers]
-nd = "nd"
-
-[tool.typos.type.css.extend-words]
-
-[tool.typos.type.man]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.man.extend-identifiers]
-Nd = "Nd"
-
-[tool.typos.type.man.extend-words]
-
-[tool.typos.type.cert]
-extend-glob = []
-check-file = false
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.cert.extend-identifiers]
-
-[tool.typos.type.cert.extend-words]
-
-[tool.typos.type.sh]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.sh.extend-identifiers]
-ot = "ot"
-
-[tool.typos.type.sh.extend-words]
-
-[tool.typos.type.vimscript]
-extend-glob = []
-extend-ignore-identifiers-re = []
-extend-ignore-words-re = []
-extend-ignore-re = []
-
-[tool.typos.type.vimscript.extend-identifiers]
-windo = "windo"
-
-[tool.typos.type.vimscript.extend-words]

requirements/common.txt

@@ -7,7 +7,7 @@ requests >= 2.26.0
 tqdm
 blake3
 py-cpuinfo
-transformers >= 4.53.2
+transformers >= 4.51.1
 huggingface-hub[hf_xet] >= 0.33.0  # Required for Xet downloads.
 tokenizers >= 0.21.1  # Required for fast incremental detokenization.
 protobuf # Required by LlamaTokenizer.
@@ -25,7 +25,7 @@ outlines_core == 0.2.10
 # required for outlines backend disk cache
 diskcache == 5.6.3
 lark == 1.2.2
-xgrammar == 0.1.21; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
+xgrammar == 0.1.19; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs
@@ -33,7 +33,7 @@ pyzmq >= 25.0.0
 msgspec
 gguf >= 0.13.0
 importlib_metadata; python_version < '3.10'
-mistral_common[opencv] >= 1.8.0
+mistral_common[opencv] >= 1.6.2
 opencv-python-headless >= 4.11.0    # required for video IO
 pyyaml
 six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that needs to be the latest version for python 3.12
@@ -47,4 +47,3 @@ python-json-logger # Used by logging as per examples/others/logging_configuratio
 scipy # Required for phi-4-multimodal-instruct
 ninja # Required for xgrammar, rocm, tpu, xpu
 pybase64 # fast base64 implementation
-cbor2 # Required for cross-language serialization of hashable objects

requirements/docs.txt

@@ -4,7 +4,6 @@ mkdocs-material
 mkdocstrings-python
 mkdocs-gen-files
 mkdocs-awesome-nav
-mkdocs-glightbox
 python-markdown-math
 regex
 ruff
@@ -12,12 +11,10 @@ ruff
 # Required for argparse hook only
 -f https://download.pytorch.org/whl/cpu
 cachetools
-cbor2
 cloudpickle
 fastapi
 msgspec
 openai
-partial-json-parser
 pillow
 psutil
 pybase64

requirements/nightly_torch_test.txt

@@ -23,7 +23,7 @@ jiwer # required for audio tests
 timm # required for internvl test
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
-mistral_common[opencv] >= 1.8.0 # required for voxtral test
+mistral_common[opencv] >= 1.6.2 # required for pixtral test
 num2words # required for smolvlm test
 opencv-python-headless >= 4.11.0 # required for video test
 datamodel_code_generator # required for minicpm3 test

requirements/test.in

@@ -28,13 +28,13 @@ torchvision==0.22.0
 transformers_stream_generator # required for qwen-vl test
 mamba_ssm # required for plamo2 test
 matplotlib # required for qwen-vl test
-mistral_common[opencv] >= 1.8.0 # required for voxtral test
+mistral_common[opencv] >= 1.6.2 # required for pixtral test
 num2words # required for smolvlm test
 opencv-python-headless >= 4.11.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]==0.4.8 # required for model evaluation test
 mteb[bm25s]>=1.38.11, <2 # required for mteb test
-transformers==4.53.2
+transformers==4.52.4
 tokenizers==0.21.1
 huggingface-hub[hf_xet]>=0.33.0  # Required for Xet downloads.
 schemathesis>=3.39.15 # Required for openai schema test.

requirements/test.txt

@@ -305,7 +305,7 @@ mbstrdecoder==1.1.3
     #   typepy
 mdurl==0.1.2
     # via markdown-it-py
-mistral-common==1.8.0
+mistral-common==1.6.2
     # via -r requirements/test.in
 more-itertools==10.5.0
     # via lm-eval
@@ -518,8 +518,6 @@ pyasn1-modules==0.4.2
     # via google-auth
 pybind11==2.13.6
     # via lm-eval
-pycountry==24.6.1
-    # via pydantic-extra-types
 pycparser==2.22
     # via cffi
 pycryptodomex==3.22.0
@@ -530,12 +528,9 @@ pydantic==2.11.5
     #   datamodel-code-generator
     #   mistral-common
     #   mteb
-    #   pydantic-extra-types
     #   ray
 pydantic-core==2.33.2
     # via pydantic
-pydantic-extra-types==2.10.5
-    # via mistral-common
 pygments==2.18.0
     # via rich
 pyparsing==3.2.0
@@ -805,7 +800,7 @@ tqdm==4.66.6
     #   transformers
 tqdm-multiprocess==0.0.11
     # via lm-eval
-transformers==4.53.2
+transformers==4.52.4
     # via
     #   -r requirements/test.in
     #   genai-perf
@@ -840,7 +835,6 @@ typing-extensions==4.12.2
     #   pqdm
     #   pydantic
     #   pydantic-core
-    #   pydantic-extra-types
     #   torch
     #   typer
     #   typing-inspection

requirements/tpu.txt

@@ -18,9 +18,9 @@ setuptools==78.1.0
 --find-links https://storage.googleapis.com/libtpu-releases/index.html
 --find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html
 --find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
-torch==2.9.0.dev20250711
-torchvision==0.24.0.dev20250711
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.9.0.dev20250711-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.9.0.dev20250711-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.9.0.dev20250711-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
+torch==2.9.0.dev20250703
+torchvision==0.24.0.dev20250703
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250703-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250703-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250703-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
 

setup.py

@@ -692,8 +692,7 @@ setup(
         "tensorizer": ["tensorizer==2.10.1"],
         "fastsafetensors": ["fastsafetensors >= 0.1.10"],
         "runai": ["runai-model-streamer", "runai-model-streamer-s3", "boto3"],
-        "audio": ["librosa", "soundfile",
-                  "mistral_common[audio]"],  # Required for audio processing
+        "audio": ["librosa", "soundfile"],  # Required for audio processing
         "video": []  # Kept for backwards compatibility
     },
     cmdclass=cmdclass,

tests/async_engine/test_api_server.py

@@ -29,7 +29,7 @@ def _query_server_long(prompt: str) -> dict:
 
 
 @pytest.fixture
-def api_server(distributed_executor_backend: str):
+def api_server(tokenizer_pool_size: int, distributed_executor_backend: str):
     script_path = Path(__file__).parent.joinpath(
         "api_server_async_engine.py").absolute()
     commands = [
@@ -40,6 +40,8 @@ def api_server(distributed_executor_backend: str):
         "facebook/opt-125m",
         "--host",
         "127.0.0.1",
+        "--tokenizer-pool-size",
+        str(tokenizer_pool_size),
         "--distributed-executor-backend",
         distributed_executor_backend,
     ]
@@ -52,8 +54,10 @@ def api_server(distributed_executor_backend: str):
     uvicorn_process.terminate()
 
 
+@pytest.mark.parametrize("tokenizer_pool_size", [0, 2])
 @pytest.mark.parametrize("distributed_executor_backend", ["mp", "ray"])
-def test_api_server(api_server, distributed_executor_backend: str):
+def test_api_server(api_server, tokenizer_pool_size: int,
+                    distributed_executor_backend: str):
     """
     Run the API server and test it.
 

tests/compile/test_config.py

@@ -26,30 +26,6 @@ def test_use_cudagraphs_dynamic(monkeypatch):
     assert not vllm_config.compilation_config.use_cudagraph
 
 
-# NB: We don't test VLLM_DISABLE_COMPILE_CACHE=0 because that depends
-# on the state of the cache directory on the current machine, which
-# may be influenced by other tests.
-@pytest.mark.parametrize("val", ["1"])
-def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
-    assert vllm.envs.VLLM_USE_V1
-
-    # spawn means that the counters are in the same process.
-    monkeypatch.setenv('VLLM_WORKER_MULTIPROC_METHOD', "spawn")
-    monkeypatch.setenv('VLLM_DISABLE_COMPILE_CACHE', val)
-
-    compilation_config = {
-        "use_cudagraph": False,  # speed things up a bit
-    }
-    with (
-            compilation_counter.expect(num_cache_entries_updated=0,
-                                       num_compiled_artifacts_saved=0),
-            # loading the model causes compilation (if enabled) to happen
-            vllm_runner('facebook/opt-125m',
-                        compilation_config=compilation_config,
-                        gpu_memory_utilization=0.4) as _):
-        pass
-
-
 @pytest.mark.parametrize("enabled", [True, False])
 def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
     assert vllm.envs.VLLM_USE_V1

tests/compile/test_fusion_all_reduce.py

@@ -1,150 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from importlib.util import find_spec
-
-import pytest
-import torch
-
-import vllm.envs as envs
-from vllm.compilation.collective_fusion import AllReduceFusionPass
-from vllm.config import (CompilationConfig, CompilationLevel, DeviceConfig,
-                         ModelConfig, PassConfig, VllmConfig)
-from vllm.distributed import tensor_model_parallel_all_reduce
-from vllm.distributed.parallel_state import (init_distributed_environment,
-                                             initialize_model_parallel)
-from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.platforms import current_platform
-from vllm.utils import update_environment_variables
-
-from ..utils import multi_gpu_test
-from .backend import TestBackend
-
-
-class TestAllReduceRMSNormModel(torch.nn.Module):
-
-    def __init__(self, hidden_size=16, eps=1e-6):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.eps = eps
-        self.norm = RMSNorm(hidden_size, eps)
-
-    def forward(self, hidden_states, residual):
-        view = hidden_states.reshape(-1, self.hidden_size)
-        all_reduce = tensor_model_parallel_all_reduce(view)
-        norm = self.norm(all_reduce)
-        return norm
-
-    def ops_in_model_before(self):
-        return [torch.ops.vllm.all_reduce.default]
-
-    def ops_in_model_after(self):
-        return [torch.ops.vllm.flashinfer_trtllm_fused_allreduce_norm.default]
-
-
-class TestAllReduceFusedAddRMSNormModel(torch.nn.Module):
-
-    def __init__(self, hidden_size=16, eps=1e-6):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.eps = eps
-        self.norm = RMSNorm(hidden_size, eps)
-
-    def forward(self, hidden_states, residual):
-        view = hidden_states.reshape(-1, self.hidden_size)
-        all_reduce = tensor_model_parallel_all_reduce(view)
-        norm, _ = self.norm(all_reduce, residual)
-        return norm
-
-    def ops_in_model_before(self):
-        return [torch.ops.vllm.all_reduce.default]
-
-    def ops_in_model_after(self):
-        return [torch.ops.vllm.flashinfer_trtllm_fused_allreduce_norm.default]
-
-
-@multi_gpu_test(num_gpus=2)
-@pytest.mark.parametrize(
-    "test_model",
-    [TestAllReduceRMSNormModel, TestAllReduceFusedAddRMSNormModel])
-@pytest.mark.parametrize("batch_size", [8])
-@pytest.mark.parametrize("seq_len", [8])
-@pytest.mark.parametrize("hidden_size", [4096])
-@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
-@pytest.mark.skipif(envs.VLLM_TARGET_DEVICE not in ["cuda"],
-                    reason="Only test on CUDA")
-@pytest.mark.skipif(not find_spec("flashinfer"),
-                    reason="flashinfer is not installed")
-@pytest.mark.skipif(not current_platform.is_device_capability(100),
-                    reason="Only test on SM100")
-def test_all_reduce_fusion_pass_replace(test_model: torch.nn.Module,
-                                        batch_size: int, seq_len: int,
-                                        hidden_size: int, dtype: torch.dtype):
-    num_processes = 2
-
-    def run_torch_spawn(fn, nprocs):
-        torch.multiprocessing.spawn(fn,
-                                    args=(num_processes, test_model,
-                                          batch_size, seq_len, hidden_size,
-                                          dtype),
-                                    nprocs=nprocs)
-
-    run_torch_spawn(all_reduce_fusion_pass_on_test_model, num_processes)
-
-
-def all_reduce_fusion_pass_on_test_model(local_rank: int, world_size: int,
-                                         test_model_cls: torch.nn.Module,
-                                         batch_size: int, seq_len: int,
-                                         hidden_size: int, dtype: torch.dtype):
-    current_platform.seed_everything(0)
-
-    device = torch.device(f"cuda:{local_rank}")
-    torch.cuda.set_device(device)
-    torch.set_default_device(device)
-    torch.set_default_dtype(dtype)
-
-    update_environment_variables({
-        'RANK': str(local_rank),
-        'LOCAL_RANK': str(local_rank),
-        'WORLD_SIZE': str(world_size),
-        'MASTER_ADDR': 'localhost',
-        'MASTER_PORT': '12345',
-    })
-
-    init_distributed_environment()
-    initialize_model_parallel(tensor_model_parallel_size=world_size)
-
-    vllm_config = VllmConfig(
-        compilation_config=CompilationConfig(level=CompilationLevel.PIECEWISE,
-                                             custom_ops=["+rms_norm"],
-                                             compile_sizes=[2, 4, 8]))
-    vllm_config.compilation_config.pass_config = PassConfig(
-        enable_fi_allreduce_fusion=True)
-    vllm_config.device_config = DeviceConfig(device=torch.device("cuda"))
-
-    # this is a fake model name to construct the model config
-    # in the vllm_config, it's not really used.
-    model_name = "nm-testing/TinyLlama-1.1B-Chat-v1.0-FP8-e2e"
-    vllm_config.model_config = ModelConfig(model=model_name,
-                                           task="auto",
-                                           tokenizer=model_name,
-                                           tokenizer_mode="auto",
-                                           trust_remote_code=True,
-                                           dtype=dtype,
-                                           seed=42)
-
-    all_reduce_fusion_pass = AllReduceFusionPass(vllm_config)
-    backend = TestBackend(all_reduce_fusion_pass)
-
-    model = test_model_cls(hidden_size)
-
-    hidden_states = torch.randn((batch_size * seq_len, hidden_size),
-                                requires_grad=False)
-    residual = torch.randn((batch_size * seq_len, hidden_size),
-                           requires_grad=False)
-
-    compiled_model = torch.compile(model, backend=backend)
-    compiled_model(hidden_states, residual)
-
-    backend.check_before_ops(model.ops_in_model_before(), fully_replaced=False)
-    backend.check_after_ops(model.ops_in_model_after())
-    del all_reduce_fusion_pass

tests/conftest.py

@@ -804,7 +804,7 @@ class VllmRunner:
 
     def get_inputs(
         self,
-        prompts: Union[list[str], list[torch.Tensor], list[int]],
+        prompts: Union[list[str], list[torch.Tensor]],
         images: Optional[PromptImageInput] = None,
         videos: Optional[PromptVideoInput] = None,
         audios: Optional[PromptAudioInput] = None,
@@ -826,16 +826,11 @@ class VllmRunner:
             if audios is not None and (audio := audios[i]) is not None:
                 multi_modal_data["audio"] = audio
 
-            text_prompt_kwargs: dict[str, Any] = {
+            text_prompt_kwargs = {
+                ("prompt" if isinstance(prompt, str) else "prompt_embeds"):
+                prompt,
                 "multi_modal_data": multi_modal_data or None
             }
-            if isinstance(prompt, str):
-                text_prompt_kwargs["prompt"] = prompt
-            elif isinstance(prompt, list):
-                text_prompt_kwargs["prompt_token_ids"] = prompt
-            else:
-                text_prompt_kwargs["prompt_embeds"] = prompt
-
             inputs.append(TextPrompt(**text_prompt_kwargs))
 
         return inputs

tests/distributed/test_pipeline_parallel.py

@@ -14,9 +14,8 @@ from typing import Literal, NamedTuple, Optional
 
 import pytest
 
-from vllm.config import _FLOAT16_NOT_SUPPORTED_MODELS, TaskOption
+from vllm.config import TaskOption
 from vllm.logger import init_logger
-from vllm.transformers_utils.config import get_config
 
 from ..models.registry import HF_EXAMPLE_MODELS
 from ..utils import compare_two_settings, create_new_process_for_each_test
@@ -159,7 +158,7 @@ TEXT_GENERATION_MODELS = {
     "databricks/dbrx-instruct": PPTestSettings.fast(load_format="dummy"),
     "Deci/DeciLM-7B-instruct": PPTestSettings.fast(),
     "deepseek-ai/deepseek-llm-7b-chat": PPTestSettings.fast(),
-    "deepseek-ai/DeepSeek-V2-Lite-Chat": PPTestSettings.fast(tp_base=2),
+    "deepseek-ai/DeepSeek-V2-Lite-Chat": PPTestSettings.fast(),
     "LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct": PPTestSettings.fast(),
     "tiiuae/falcon-7b": PPTestSettings.fast(),
     "google/gemma-1.1-2b-it": PPTestSettings.fast(),
@@ -211,11 +210,9 @@ TEXT_GENERATION_MODELS = {
 
 EMBEDDING_MODELS = {  # type: ignore[var-annotated]
     # [Text-only]
-    "intfloat/e5-mistral-7b-instruct": PPTestSettings.fast(task="embed"),
-    "BAAI/bge-multilingual-gemma2": PPTestSettings.fast(task="embed"),
-    "Qwen/Qwen2.5-Math-RM-72B": PPTestSettings.fast(
-        load_format="dummy", task="embed"
-    ),
+    "intfloat/e5-mistral-7b-instruct": PPTestSettings.fast(),
+    "BAAI/bge-multilingual-gemma2": PPTestSettings.fast(),
+    "Qwen/Qwen2.5-Math-RM-72B": PPTestSettings.fast(load_format="dummy"),
 }
 
 MULTIMODAL_MODELS = {
@@ -251,7 +248,6 @@ TEST_MODELS = [
     "meta-llama/Llama-3.2-1B-Instruct",
     "ArthurZ/Ilama-3.2-1B",
     "ibm/PowerLM-3b",
-    "deepseek-ai/DeepSeek-V2-Lite-Chat",
     # [LANGUAGE EMBEDDING]
     "intfloat/e5-mistral-7b-instruct",
     "BAAI/bge-multilingual-gemma2",
@@ -291,11 +287,6 @@ def _compare_tp(
     trust_remote_code = model_info.trust_remote_code
     tokenizer_mode = model_info.tokenizer_mode
     hf_overrides = model_info.hf_overrides
-    hf_config = get_config(model_id, trust_remote_code)
-
-    dtype = "float16"
-    if hf_config.model_type in _FLOAT16_NOT_SUPPORTED_MODELS:
-        dtype = "bfloat16"
 
     if load_format == "dummy":
         # Avoid OOM
@@ -325,7 +316,7 @@ def _compare_tp(
     common_args = [
         # use half precision for speed and memory savings in CI environment
         "--dtype",
-        dtype,
+        "float16",
         "--max-model-len",
         "2048",
         "--max-num-seqs",
@@ -347,7 +338,6 @@ def _compare_tp(
         common_args.extend(["--hf-overrides", json.dumps(hf_overrides)])
 
     specific_case = tp_size == 2 and pp_size == 2 and chunked_prefill
-    testing_ray_compiled_graph = False
     if distributed_backend == "ray" and (vllm_major_version == "1"
                                          or specific_case):
         # For V1, test Ray Compiled Graph for all the tests
@@ -361,7 +351,6 @@ def _compare_tp(
         # Temporary. Currently when zeromq + SPMD is used, it does not properly
         # terminate because of a Ray Compiled Graph issue.
         common_args.append("--disable-frontend-multiprocessing")
-        testing_ray_compiled_graph = True
     elif distributed_backend == "mp":
         # Both V0/V1 of multiprocessing executor support PP
         pp_env = {
@@ -405,6 +394,7 @@ def _compare_tp(
                              tp_env,
                              method=method)
     except Exception:
+        testing_ray_compiled_graph = pp_env is not None
         if testing_ray_compiled_graph and vllm_major_version == "0":
             # Ray Compiled Graph tests are flaky for V0,
             # so we don't want to fail the test

tests/distributed/test_pynccl.py

@@ -4,7 +4,6 @@
 import multiprocessing
 import os
 
-import numpy as np
 import pytest
 import torch
 import torch.distributed
@@ -178,38 +177,6 @@ def test_pynccl_all_gather():
     distributed_run(all_gather_worker_fn, 2)
 
 
-@worker_fn_wrapper
-def all_gatherv_worker_fn():
-    pynccl_comm = PyNcclCommunicator(get_world_group().cpu_group,
-                                     device=get_world_group().device)
-
-    rank = pynccl_comm.rank
-    world_size = pynccl_comm.world_size
-    device = f'cuda:{pynccl_comm.rank}'
-
-    assert world_size <= 8
-    sizes = [81, 20, 57, 52, 81, 5, 49, 49][:world_size]
-    num_elems = sizes[rank]
-    tensor = torch.arange(num_elems, dtype=torch.float32,
-                          device=device) + rank * 100
-    result = torch.zeros(sum(sizes), dtype=torch.float32, device=device)
-
-    expected = torch.cat([
-        torch.arange(sizes[r], dtype=torch.float32) + r * 100
-        for r in range(world_size)
-    ]).to(device)
-
-    pynccl_comm.all_gatherv(result, tensor, sizes=sizes)
-    torch.cuda.synchronize()
-    torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8)
-
-
-@pytest.mark.skipif(torch.cuda.device_count() < 2,
-                    reason="Need at least 2 GPUs to run the test.")
-def test_pynccl_all_gatherv():
-    distributed_run(all_gatherv_worker_fn, 2)
-
-
 @worker_fn_wrapper
 def reduce_scatter_worker_fn():
     pynccl_comm = PyNcclCommunicator(get_world_group().cpu_group,
@@ -247,43 +214,6 @@ def test_pynccl_reduce_scatter():
     distributed_run(reduce_scatter_worker_fn, 2)
 
 
-@worker_fn_wrapper
-def reduce_scatterv_worker_fn():
-    pynccl_comm = PyNcclCommunicator(get_world_group().cpu_group,
-                                     device=get_world_group().device)
-
-    rank = pynccl_comm.rank
-    world_size = pynccl_comm.world_size
-    device = f'cuda:{pynccl_comm.rank}'
-
-    assert world_size <= 8
-    sizes = [81, 20, 57, 52, 81, 5, 49, 49][:world_size]
-    num_elems = sum(sizes)
-    tensor = torch.arange(num_elems, dtype=torch.float32,
-                          device=device) + rank * 100
-    result = torch.zeros(sizes[rank], dtype=torch.float32, device=device)
-
-    # Calculate expected result for this rank's chunk
-    all_tensors = [
-        torch.arange(num_elems, dtype=torch.float32) + r * 100
-        for r in range(world_size)
-    ]
-    sizes_cumsum = np.cumsum(sizes)
-    start = 0 if rank == 0 else sizes_cumsum[rank - 1]
-    end = sizes_cumsum[rank]
-    expected = sum(tensor[start:end] for tensor in all_tensors).to(device)
-
-    pynccl_comm.reduce_scatterv(result, tensor, sizes=sizes)
-    torch.cuda.synchronize()
-    torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-8)
-
-
-@pytest.mark.skipif(torch.cuda.device_count() < 2,
-                    reason="Need at least 2 GPUs to run the test.")
-def test_pynccl_reduce_scatterv():
-    distributed_run(reduce_scatterv_worker_fn, 2)
-
-
 @pytest.mark.skipif(torch.cuda.device_count() < 2,
                     reason="Need at least 2 GPUs to run the test.")
 def test_pynccl_with_cudagraph():

tests/engine/test_arg_utils.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
-from argparse import ArgumentError
+from argparse import ArgumentError, ArgumentTypeError
 from contextlib import nullcontext
 from dataclasses import dataclass, field
 from typing import Annotated, Literal, Optional
@@ -12,8 +12,8 @@ import pytest
 from vllm.config import CompilationConfig, config
 from vllm.engine.arg_utils import (EngineArgs, contains_type, get_kwargs,
                                    get_type, get_type_hints, is_not_builtin,
-                                   is_type, literal_to_kwargs, optional_type,
-                                   parse_type)
+                                   is_type, literal_to_kwargs, nullable_kvs,
+                                   optional_type, parse_type)
 from vllm.utils import FlexibleArgumentParser
 
 
@@ -25,10 +25,18 @@ from vllm.utils import FlexibleArgumentParser
         "foo": 1,
         "bar": 2
     }),
+    (json.loads, "foo=1,bar=2", {
+        "foo": 1,
+        "bar": 2
+    }),
 ])
 def test_parse_type(type, value, expected):
     parse_type_func = parse_type(type)
-    assert parse_type_func(value) == expected
+    context = nullcontext()
+    if value == "foo=1,bar=2":
+        context = pytest.warns(DeprecationWarning)
+    with context:
+        assert parse_type_func(value) == expected
 
 
 def test_optional_type():
@@ -195,6 +203,34 @@ def test_get_kwargs():
     assert kwargs["from_cli_config2"]["type"]('{"field": 2}').field == 4
 
 
+@pytest.mark.parametrize(("arg", "expected"), [
+    (None, dict()),
+    ("image=16", {
+        "image": 16
+    }),
+    ("image=16,video=2", {
+        "image": 16,
+        "video": 2
+    }),
+    ("Image=16, Video=2", {
+        "image": 16,
+        "video": 2
+    }),
+])
+def test_limit_mm_per_prompt_parser(arg, expected):
+    """This functionality is deprecated and will be removed in the future.
+    This argument should be passed as JSON string instead.
+    
+    TODO: Remove with nullable_kvs."""
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--limit-mm-per-prompt", arg])
+
+    assert args.limit_mm_per_prompt == expected
+
+
 @pytest.mark.parametrize(
     ("arg", "expected"),
     [
@@ -290,6 +326,18 @@ def test_prefix_cache_default():
     assert not engine_args.enable_prefix_caching
 
 
+@pytest.mark.parametrize(
+    ("arg"),
+    [
+        "image",  # Missing =
+        "image=4,image=5",  # Conflicting values
+        "image=video=4"  # Too many = in tokenized arg
+    ])
+def test_bad_nullable_kvs(arg):
+    with pytest.raises(ArgumentTypeError):
+        nullable_kvs(arg)
+
+
 # yapf: disable
 @pytest.mark.parametrize(("arg", "expected", "option"), [
     (None, None, "mm-processor-kwargs"),

tests/entrypoints/llm/test_accuracy.py

@@ -71,8 +71,9 @@ def test_lm_eval_accuracy_v1_engine(model, monkeypatch: pytest.MonkeyPatch):
             # Limit compilation time for TPU V1
 
             if model == "google/gemma-3-1b-it":
-                # TPU + google/gemma-3-1b-it + xet doesn't work well.
-                m.setenv("HF_HUB_DISABLE_XET", "1")
+                pytest.skip(
+                    "Temporarily disabled due to test failures"
+                    "(timeout or accuracy mismatch). Re-enable once fixed.")
 
             more_args = "max_model_len=2048,max_num_seqs=64"
 

tests/entrypoints/openai/test_classification.py

@@ -176,8 +176,4 @@ async def test_invocations(server: RemoteOpenAIServer):
     invocation_output = invocation_response.json()
 
     assert classification_output.keys() == invocation_output.keys()
-    for classification_data, invocation_data in zip(
-            classification_output["data"], invocation_output["data"]):
-        assert classification_data.keys() == invocation_data.keys()
-        assert classification_data["probs"] == pytest.approx(
-            invocation_data["probs"], rel=0.01)
+    assert classification_output["data"] == invocation_output["data"]

tests/entrypoints/openai/test_embedding.py

@@ -14,7 +14,6 @@ from vllm.transformers_utils.tokenizer import get_tokenizer
 
 from ...models.language.pooling.embed_utils import (
     run_embedding_correctness_test)
-from ...models.utils import check_embeddings_close
 from ...utils import RemoteOpenAIServer
 
 MODEL_NAME = "intfloat/multilingual-e5-small"
@@ -322,13 +321,7 @@ async def test_invocations(server: RemoteOpenAIServer,
     invocation_output = invocation_response.json()
 
     assert completion_output.keys() == invocation_output.keys()
-    for completion_data, invocation_data in zip(completion_output["data"],
-                                                invocation_output["data"]):
-        assert completion_data.keys() == invocation_data.keys()
-        check_embeddings_close(embeddings_0_lst=[completion_data["embedding"]],
-                               embeddings_1_lst=[invocation_data["embedding"]],
-                               name_0="completion",
-                               name_1="invocation")
+    assert completion_output["data"] == invocation_output["data"]
 
 
 @pytest.mark.asyncio
@@ -362,10 +355,4 @@ async def test_invocations_conversation(server: RemoteOpenAIServer):
     invocation_output = invocation_response.json()
 
     assert chat_output.keys() == invocation_output.keys()
-    for chat_data, invocation_data in zip(chat_output["data"],
-                                          invocation_output["data"]):
-        assert chat_data.keys() == invocation_data.keys()
-        check_embeddings_close(embeddings_0_lst=[chat_data["embedding"]],
-                               embeddings_1_lst=[invocation_data["embedding"]],
-                               name_0="chat",
-                               name_1="invocation")
+    assert chat_output["data"] == invocation_output["data"]

tests/entrypoints/openai/test_openai_schema.py

@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import json
 from typing import Final
 
 import pytest
@@ -30,7 +29,7 @@ def server():
         "--enforce-eager",
         "--trust-remote-code",
         "--limit-mm-per-prompt",
-        json.dumps({"image": MAXIMUM_IMAGES}),
+        f"image={MAXIMUM_IMAGES}",
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:

tests/entrypoints/openai/test_pooling.py

@@ -281,13 +281,7 @@ async def test_invocations(server: RemoteOpenAIServer):
     invocation_output = invocation_response.json()
 
     assert completion_output.keys() == invocation_output.keys()
-    for completion_data, invocation_data in zip(completion_output["data"],
-                                                invocation_output["data"]):
-        assert completion_data.keys() == invocation_data.keys()
-        check_embeddings_close(embeddings_0_lst=completion_data["data"],
-                               embeddings_1_lst=invocation_data["data"],
-                               name_0="completion",
-                               name_1="invocation")
+    assert completion_output["data"] == invocation_output["data"]
 
 
 @pytest.mark.asyncio
@@ -320,10 +314,4 @@ async def test_invocations_conversation(server: RemoteOpenAIServer):
     invocation_output = invocation_response.json()
 
     assert chat_output.keys() == invocation_output.keys()
-    for chat_data, invocation_data in zip(chat_output["data"],
-                                          invocation_output["data"]):
-        assert chat_data.keys() == invocation_data.keys()
-        check_embeddings_close(embeddings_0_lst=chat_data["data"],
-                               embeddings_1_lst=invocation_data["data"],
-                               name_0="chat",
-                               name_1="invocation")
+    assert chat_output["data"] == invocation_output["data"]

tests/entrypoints/openai/test_rerank.py

@@ -120,8 +120,4 @@ def test_invocations(server: RemoteOpenAIServer):
     invocation_output = invocation_response.json()
 
     assert rerank_output.keys() == invocation_output.keys()
-    for rerank_result, invocations_result in zip(rerank_output["results"],
-                                                 invocation_output["results"]):
-        assert rerank_result.keys() == invocations_result.keys()
-        assert rerank_result["relevance_score"] == pytest.approx(
-            invocations_result["relevance_score"], rel=0.01)
+    assert rerank_output["results"] == invocation_output["results"]

tests/entrypoints/openai/test_score.py

@@ -215,8 +215,4 @@ class TestModel:
         invocation_output = invocation_response.json()
 
         assert score_output.keys() == invocation_output.keys()
-        for score_data, invocation_data in zip(score_output["data"],
-                                               invocation_output["data"]):
-            assert score_data.keys() == invocation_data.keys()
-            assert score_data["score"] == pytest.approx(
-                invocation_data["score"], rel=0.01)
+        assert score_output["data"] == invocation_output["data"]

tests/entrypoints/openai/test_tokenization.py

@@ -32,7 +32,6 @@ def server(zephyr_lora_added_tokens_files: str):  # noqa: F811
         f"zephyr-lora2={zephyr_lora_added_tokens_files}",
         "--max-lora-rank",
         "64",
-        "--enable-tokenizer-info-endpoint",
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
@@ -284,106 +283,3 @@ async def test_detokenize(
     response.raise_for_status()
 
     assert response.json() == {"prompt": prompt}
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize(
-    "model_name,tokenizer_name",
-    [(MODEL_NAME, MODEL_NAME), ("zephyr-lora2", "zephyr-lora2")],
-    indirect=["tokenizer_name"],
-)
-async def test_tokenizer_info_basic(
-    server: RemoteOpenAIServer,
-    model_name: str,
-    tokenizer_name: str,
-):
-    """Test basic tokenizer info endpoint functionality."""
-    response = requests.get(server.url_for("tokenizer_info"))
-    response.raise_for_status()
-    result = response.json()
-    assert "tokenizer_class" in result
-    assert isinstance(result["tokenizer_class"], str)
-    assert result["tokenizer_class"]
-
-
-@pytest.mark.asyncio
-async def test_tokenizer_info_schema(server: RemoteOpenAIServer):
-    """Test that the response matches expected schema types."""
-    response = requests.get(server.url_for("tokenizer_info"))
-    response.raise_for_status()
-    result = response.json()
-    field_types = {
-        "add_bos_token": bool,
-        "add_prefix_space": bool,
-        "clean_up_tokenization_spaces": bool,
-        "split_special_tokens": bool,
-        "bos_token": str,
-        "eos_token": str,
-        "pad_token": str,
-        "unk_token": str,
-        "chat_template": str,
-        "errors": str,
-        "model_max_length": int,
-        "additional_special_tokens": list,
-        "added_tokens_decoder": dict,
-    }
-    for field, expected_type in field_types.items():
-        if field in result and result[field] is not None:
-            assert isinstance(
-                result[field],
-                expected_type), (f"{field} should be {expected_type.__name__}")
-
-
-@pytest.mark.asyncio
-async def test_tokenizer_info_added_tokens_structure(
-    server: RemoteOpenAIServer, ):
-    """Test added_tokens_decoder structure if present."""
-    response = requests.get(server.url_for("tokenizer_info"))
-    response.raise_for_status()
-    result = response.json()
-    added_tokens = result.get("added_tokens_decoder")
-    if added_tokens:
-        for token_id, token_info in added_tokens.items():
-            assert isinstance(token_id, str), "Token IDs should be strings"
-            assert isinstance(token_info, dict), "Token info should be a dict"
-            assert "content" in token_info, "Token info should have content"
-            assert "special" in token_info, (
-                "Token info should have special flag")
-            assert isinstance(token_info["special"],
-                              bool), ("Special flag should be boolean")
-
-
-@pytest.mark.asyncio
-async def test_tokenizer_info_consistency_with_tokenize(
-    server: RemoteOpenAIServer, ):
-    """Test that tokenizer info is consistent with tokenization endpoint."""
-    info_response = requests.get(server.url_for("tokenizer_info"))
-    info_response.raise_for_status()
-    info = info_response.json()
-    tokenize_response = requests.post(
-        server.url_for("tokenize"),
-        json={
-            "model": MODEL_NAME,
-            "prompt": "Hello world!"
-        },
-    )
-    tokenize_response.raise_for_status()
-    tokenize_result = tokenize_response.json()
-    info_max_len = info.get("model_max_length")
-    tokenize_max_len = tokenize_result.get("max_model_len")
-    if info_max_len and tokenize_max_len:
-        assert info_max_len >= tokenize_max_len, (
-            "Info max length should be >= tokenize max length")
-
-
-@pytest.mark.asyncio
-async def test_tokenizer_info_chat_template(server: RemoteOpenAIServer):
-    """Test chat template is properly included."""
-    response = requests.get(server.url_for("tokenizer_info"))
-    response.raise_for_status()
-    result = response.json()
-    chat_template = result.get("chat_template")
-    if chat_template:
-        assert isinstance(chat_template,
-                          str), ("Chat template should be a string")
-        assert chat_template.strip(), "Chat template should not be empty"

tests/entrypoints/openai/test_transcription_validation.py

@@ -17,11 +17,6 @@ from vllm.assets.audio import AudioAsset
 
 from ...utils import RemoteOpenAIServer
 
-MISTRAL_FORMAT_ARGS = [
-    "--tokenizer_mode", "mistral", "--config_format", "mistral",
-    "--load_format", "mistral"
-]
-
 
 @pytest.fixture
 def mary_had_lamb():
@@ -38,15 +33,9 @@ def winning_call():
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize(
-    "model_name",
-    ["openai/whisper-large-v3-turbo", "mistralai/Voxtral-Mini-3B-2507"])
-async def test_basic_audio(mary_had_lamb, model_name):
+async def test_basic_audio(mary_had_lamb):
+    model_name = "openai/whisper-large-v3-turbo"
     server_args = ["--enforce-eager"]
-
-    if model_name.startswith("mistralai"):
-        server_args += MISTRAL_FORMAT_ARGS
-
     # Based on https://github.com/openai/openai-cookbook/blob/main/examples/Whisper_prompting_guide.ipynb.
     with RemoteOpenAIServer(model_name, server_args) as remote_server:
         client = remote_server.get_async_client()
@@ -76,13 +65,10 @@ async def test_bad_requests(mary_had_lamb):
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", ["openai/whisper-large-v3-turbo"])
-async def test_long_audio_request(mary_had_lamb, model_name):
+async def test_long_audio_request(mary_had_lamb):
+    model_name = "openai/whisper-large-v3-turbo"
     server_args = ["--enforce-eager"]
 
-    if model_name.startswith("openai"):
-        return
-
     mary_had_lamb.seek(0)
     audio, sr = librosa.load(mary_had_lamb)
     # Add small silence after each audio for repeatability in the split process
@@ -101,8 +87,7 @@ async def test_long_audio_request(mary_had_lamb, model_name):
             response_format="text",
             temperature=0.0)
         out = json.loads(transcription)['text']
-        counts = out.count("Mary had a little lamb")
-        assert counts == 10, counts
+        assert out.count("Mary had a little lamb") == 10
 
 
 @pytest.mark.asyncio
@@ -169,8 +154,7 @@ async def test_streaming_response(winning_call):
                 file=winning_call,
                 language="en",
                 temperature=0.0,
-                extra_body=dict(stream=True),
-                timeout=30)
+                extra_body=dict(stream=True))
             # Reconstruct from chunks and validate
             async for chunk in res:
                 # just a chunk
@@ -200,8 +184,7 @@ async def test_stream_options(winning_call):
                 temperature=0.0,
                 extra_body=dict(stream=True,
                                 stream_include_usage=True,
-                                stream_continuous_usage_stats=True),
-                timeout=30)
+                                stream_continuous_usage_stats=True))
             final = False
             continuous = True
             async for chunk in res:

tests/entrypoints/openai/test_translation_validation.py

@@ -39,8 +39,8 @@ async def test_basic_audio(foscolo):
             # TODO remove once language detection is implemented
             extra_body=dict(language="it"),
             temperature=0.0)
-        out = json.loads(translation)['text'].strip().lower()
-        assert "greek sea" in out
+        out = json.loads(translation)['text'].strip()
+        assert "Nor will I ever touch the sacred" in out
 
 
 @pytest.mark.asyncio
@@ -168,4 +168,5 @@ async def test_long_audio_request(foscolo):
             response_format="text",
             temperature=0.0)
         out = json.loads(translation)['text'].strip().lower()
-        assert out.count("greek sea") == 2
+        # TODO investigate higher model uncertainty in for longer translations.
+        assert out.count("nor will i ever") == 2

tests/kernels/moe/modular_kernel_tools/common.py

@@ -416,7 +416,7 @@ class RankTensors:
         # We dequant and use that as hidden_states so the tests are stable.
         # quantizing and dequantizing yield slightly different results
         # depending on the hardware. Here we, quantize and dequantize
-        # first - so further quantize and dequantize will yield the same
+        # first - so further quantize and dequantize will yeild the same
         # values.
         if config.is_per_tensor_act_quant:
             a_q, a_scales = ops.scaled_fp8_quant(

tests/kernels/moe/test_batched_moe.py

@@ -6,6 +6,7 @@ from typing import Optional
 
 import pytest
 import torch
+import triton.language as tl
 
 from tests.kernels.moe.utils import (batched_moe,
                                      make_quantized_test_activations,
@@ -17,7 +18,6 @@ from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
     invoke_moe_batched_triton_kernel)
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
 from vllm.platforms import current_platform
-from vllm.triton_utils import tl
 
 MNK_FACTORS = [
     (1, 128, 128),

tests/kernels/moe/test_deepgemm.py

@@ -13,10 +13,9 @@ import torch
 
 # vLLM fused-expert reference (Triton fallback + DeepGEMM option)
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
-from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    per_token_group_quant_fp8)
 from vllm.utils import has_deep_gemm
-from vllm.utils.deep_gemm import calc_diff, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (calc_diff, per_block_cast_to_fp8,
+                                  per_token_group_cast_to_fp8)
 
 BLOCK_SIZE = [128, 128]
 
@@ -82,7 +81,7 @@ def run_single_case(m, n, k, topk, num_experts, block_size):
     """
     tokens_bf16 = torch.randn(
         m, k, device="cuda", dtype=torch.bfloat16).clamp_min_(-1).clamp_max_(1)
-    _, a1_scale = per_token_group_quant_fp8(tokens_bf16, block_size[1])
+    _, a1_scale = per_token_group_cast_to_fp8(tokens_bf16, block_size[1])
 
     # expert weight tensors
     w1, w2, w1_s, w2_s = make_block_quant_fp8_weights(num_experts, n, k,
@@ -95,7 +94,7 @@ def run_single_case(m, n, k, topk, num_experts, block_size):
     topk_weights, topk_ids = torch.topk(router_logits, k=topk, dim=-1)
     topk_weights = torch.nn.functional.softmax(topk_weights, dim=-1)
 
-    # triton reference
+    # triton referrence
     out_triton = fused_experts(
         hidden_states=tokens_bf16,
         w1=w1,

tests/kernels/quantization/test_block_fp8.py

@@ -15,7 +15,8 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     w8a8_block_fp8_matmul)
 from vllm.platforms import current_platform
 from vllm.utils import has_deep_gemm
-from vllm.utils.deep_gemm import fp8_gemm_nt, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (fp8_gemm_nt, per_block_cast_to_fp8,
+                                  per_token_group_cast_to_fp8)
 
 if current_platform.get_device_capability() < (9, 0):
     pytest.skip("FP8 Triton requires CUDA 9.0 or higher",
@@ -116,7 +117,7 @@ def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
     A_fp32 = (torch.rand(M, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
     B_fp32 = (torch.rand(N, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
 
-    A_fp8, As_fp8 = per_token_group_quant_fp8(A_fp32, block_size[1])
+    A_fp8, As_fp8 = per_token_group_cast_to_fp8(A_fp32, block_size[1])
     B_fp8, Bs_fp8 = per_block_cast_to_fp8(B_fp32)
 
     As = As_fp8.to(torch.float32)

tests/models/language/generation/test_common.py

@@ -90,7 +90,7 @@ AITER_MODEL_LIST = [
             marks=[pytest.mark.core_model],
         ),
         pytest.param(
-            "allenai/OLMoE-1B-7B-0924-Instruct",
+            "Qwen/Qwen1.5-MoE-A2.7B-Chat",
             marks=[pytest.mark.cpu_model],
         )
     ])

tests/models/language/generation/test_hybrid.py

@@ -61,6 +61,14 @@ V1_SUPPORTED_MODELS = [
     "tiiuae/Falcon-H1-0.5B-Base",
 ]
 
+ATTN_BLOCK_SIZES = {
+    "ibm-ai-platform/Bamba-9B-v1": 528,
+    "Zyphra/Zamba2-1.2B-instruct": 80,
+    "nvidia/Nemotron-H-8B-Base-8K": 528,
+    "ibm-granite/granite-4.0-tiny-preview": 400,
+    "tiiuae/Falcon-H1-0.5B-Base": 800,
+}
+
 # Avoid OOM
 MAX_NUM_SEQS = 4
 
@@ -97,6 +105,11 @@ def test_models(
             example_prompts, max_tokens, num_logprobs)
 
     if model in V1_SUPPORTED_MODELS:
+        if model in HYBRID_MODELS and model in ATTN_BLOCK_SIZES:
+            block_size = ATTN_BLOCK_SIZES[model]
+        else:
+            block_size = 16
+
         with monkeypatch.context() as m:
             m.setenv("VLLM_USE_V1", "1")
             if model in HYBRID_MODELS:
@@ -105,7 +118,8 @@ def test_models(
             with vllm_runner(model,
                              max_num_seqs=MAX_NUM_SEQS,
                              enforce_eager=True,
-                             enable_prefix_caching=False) as vllm_model:
+                             enable_prefix_caching=False,
+                             block_size=block_size) as vllm_model:
                 vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
                     example_prompts, max_tokens, num_logprobs)
     else:

tests/models/language/pooling/mteb_utils.py

@@ -268,8 +268,7 @@ def mteb_test_rerank_models(hf_runner,
                             model_info: RerankModelInfo,
                             vllm_extra_kwargs=None,
                             hf_model_callback=None,
-                            vllm_mteb_encoder=VllmMtebEncoder,
-                            atol=MTEB_RERANK_TOL):
+                            vllm_mteb_encoder=VllmMtebEncoder):
     if not model_info.enable_test:
         # A model family has many models with the same architecture,
         # and we don't need to test each one.
@@ -302,4 +301,4 @@ def mteb_test_rerank_models(hf_runner,
     print("SentenceTransformers:", st_dtype, st_main_score)
     print("Difference:", st_main_score - vllm_main_score)
 
-    assert st_main_score == pytest.approx(vllm_main_score, abs=atol)
+    assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_RERANK_TOL)

tests/models/language/pooling/test_qwen3_reranker.py

@@ -6,7 +6,6 @@ import pytest
 import torch
 
 from tests.conftest import HfRunner
-from tests.utils import multi_gpu_test
 
 from .mteb_utils import RerankModelInfo, mteb_test_rerank_models
 
@@ -88,29 +87,3 @@ def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
 
     mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
                             vllm_extra_kwargs)
-
-
-@pytest.mark.parametrize("model_info", RERANK_MODELS)
-@multi_gpu_test(num_gpus=2)
-def test_rerank_models_mteb_tp(vllm_runner,
-                               model_info: RerankModelInfo) -> None:
-
-    assert model_info.architecture == "Qwen3ForSequenceClassification"
-
-    vllm_extra_kwargs: dict[str, Any] = {
-        "hf_overrides": {
-            "architectures": ["Qwen3ForSequenceClassification"],
-            "classifier_from_token": ["no", "yes"],
-            "is_original_qwen3_reranker": True,
-        },
-        "tensor_parallel_size": 2,
-    }
-
-    if model_info.name == "Qwen/Qwen3-Reranker-4B":
-        vllm_extra_kwargs["max_num_seqs"] = 1
-
-    mteb_test_rerank_models(Qwen3RerankerHfRunner,
-                            vllm_runner,
-                            model_info,
-                            vllm_extra_kwargs,
-                            atol=1.2e-2)

tests/models/multimodal/generation/test_common.py

@@ -318,7 +318,6 @@ VLM_TEST_SETTINGS = {
         num_logprobs=10,
         image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
         auto_cls=AutoModelForImageTextToText,
-        marks=[large_gpu_mark(min_gb=32)],
     ),
     "glm4_1v-video": VLMTestInfo(
         models=["THUDM/GLM-4.1V-9B-Thinking"],
@@ -332,7 +331,8 @@ VLM_TEST_SETTINGS = {
             inputs=custom_inputs.video_with_metadata_glm4_1v(),
             limit_mm_per_prompt={"video": 1},
         )],
-        marks=[large_gpu_mark(min_gb=32)],
+        # This is needed to run on machine with 24GB VRAM
+        vllm_runner_kwargs={"gpu_memory_utilization": 0.95},
     ),
     "h2ovl": VLMTestInfo(
         models = [

tests/models/multimodal/generation/test_voxtral.py

@@ -1,115 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import json
-
-import pytest
-import pytest_asyncio
-from mistral_common.audio import Audio
-from mistral_common.protocol.instruct.messages import (AudioChunk, RawAudio,
-                                                       TextChunk, UserMessage)
-
-from vllm.transformers_utils.tokenizer import MistralTokenizer
-
-from ....conftest import AudioTestAssets
-from ....utils import RemoteOpenAIServer
-from .test_ultravox import MULTI_AUDIO_PROMPT, run_multi_audio_test
-
-MODEL_NAME = "mistralai/Voxtral-Mini-3B-2507"
-MISTRAL_FORMAT_ARGS = [
-    "--tokenizer_mode", "mistral", "--config_format", "mistral",
-    "--load_format", "mistral"
-]
-
-
-@pytest.fixture()
-def server(request, audio_assets: AudioTestAssets):
-    args = [
-        "--enforce-eager",
-        "--limit-mm-per-prompt",
-        json.dumps({"audio": len(audio_assets)}),
-    ] + MISTRAL_FORMAT_ARGS
-
-    with RemoteOpenAIServer(MODEL_NAME,
-                            args,
-                            env_dict={"VLLM_AUDIO_FETCH_TIMEOUT":
-                                      "30"}) as remote_server:
-        yield remote_server
-
-
-@pytest_asyncio.fixture
-async def client(server):
-    async with server.get_async_client() as async_client:
-        yield async_client
-
-
-def _get_prompt(audio_assets, question):
-    tokenizer = MistralTokenizer.from_pretrained(MODEL_NAME)
-
-    audios = [
-        Audio.from_file(str(audio_assets[i].get_local_path()), strict=False)
-        for i in range(len(audio_assets))
-    ]
-    audio_chunks = [
-        AudioChunk(input_audio=RawAudio.from_audio(audio)) for audio in audios
-    ]
-
-    text_chunk = TextChunk(text=question)
-    messages = [UserMessage(content=[*audio_chunks, text_chunk]).to_openai()]
-
-    return tokenizer.apply_chat_template(messages=messages)
-
-
-@pytest.mark.core_model
-@pytest.mark.parametrize("dtype", ["half"])
-@pytest.mark.parametrize("max_tokens", [128])
-@pytest.mark.parametrize("num_logprobs", [5])
-def test_models_with_multiple_audios(vllm_runner,
-                                     audio_assets: AudioTestAssets, dtype: str,
-                                     max_tokens: int,
-                                     num_logprobs: int) -> None:
-    vllm_prompt = _get_prompt(audio_assets, MULTI_AUDIO_PROMPT)
-    run_multi_audio_test(
-        vllm_runner,
-        [(vllm_prompt, [audio.audio_and_sample_rate
-                        for audio in audio_assets])],
-        MODEL_NAME,
-        dtype=dtype,
-        max_tokens=max_tokens,
-        num_logprobs=num_logprobs,
-        tokenizer_mode="mistral",
-    )
-
-
-@pytest.mark.asyncio
-async def test_online_serving(client, audio_assets: AudioTestAssets):
-    """Exercises online serving with/without chunked prefill enabled."""
-
-    def asset_to_chunk(asset):
-        audio = Audio.from_file(str(asset.get_local_path()), strict=False)
-        audio.format = "wav"
-        audio_dict = AudioChunk.from_audio(audio).to_openai()
-        return audio_dict
-
-    audio_chunks = [asset_to_chunk(asset) for asset in audio_assets]
-    messages = [{
-        "role":
-        "user",
-        "content": [
-            *audio_chunks,
-            {
-                "type":
-                "text",
-                "text":
-                f"What's happening in these {len(audio_assets)} audio clips?"
-            },
-        ],
-    }]
-
-    chat_completion = await client.chat.completions.create(model=MODEL_NAME,
-                                                           messages=messages,
-                                                           max_tokens=10)
-
-    assert len(chat_completion.choices) == 1
-    choice = chat_completion.choices[0]
-    assert choice.finish_reason == "length"

tests/models/multimodal/pooling/test_jinavl_reranker.py

@@ -1,15 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from typing import Union
 
 import pytest
 from transformers import AutoModel
 
-from vllm.entrypoints.chat_utils import ChatCompletionContentPartImageParam
-from vllm.entrypoints.score_utils import ScoreMultiModalParam
-
-from ....conftest import HfRunner, VllmRunner
-
 model_name = "jinaai/jina-reranker-m0"
 
 mm_processor_kwargs = {
@@ -20,90 +14,73 @@ mm_processor_kwargs = {
 limit_mm_per_prompt = {"image": 2}
 
 
-def vllm_reranker(
-    vllm_runner: type[VllmRunner],
-    model_name: str,
-    dtype: str,
-    query_strs: list[str],
-    document_strs: list[str],
-    query_type: str = "text",
-    doc_type: str = "text",
-):
+def vllm_reranker(model_name,
+                  query,
+                  documents,
+                  query_type="text",
+                  doc_type="text"):
+    from vllm import LLM
 
-    def create_image_param(url: str) -> ChatCompletionContentPartImageParam:
+    model = LLM(
+        model=model_name,
+        task="score",
+        max_model_len=32768,
+        mm_processor_kwargs=mm_processor_kwargs,
+        limit_mm_per_prompt=limit_mm_per_prompt,
+    )
+
+    def create_image_param(url: str):
         return {"type": "image_url", "image_url": {"url": f"{url}"}}
 
-    query: Union[list[str], ScoreMultiModalParam]
-    if query_type == "text":
-        query = query_strs
-    elif query_type == "image":
-        query = ScoreMultiModalParam(
-            content=[create_image_param(url) for url in query_strs])
+    if query_type == "image":
+        query = {"content": [create_image_param(url) for url in query]}
 
-    documents: Union[list[str], ScoreMultiModalParam]
-    if doc_type == "text":
-        documents = document_strs
-    elif doc_type == "image":
-        documents = ScoreMultiModalParam(
-            content=[create_image_param(url) for url in document_strs])
+    if doc_type == "image":
+        documents = {"content": [create_image_param(url) for url in documents]}
 
-    with vllm_runner(
-            model_name,
-            task="score",
-            dtype=dtype,
-            max_num_seqs=2,
-            max_model_len=2048,
-            mm_processor_kwargs=mm_processor_kwargs,
-            limit_mm_per_prompt=limit_mm_per_prompt,
-    ) as vllm_model:
-        outputs = vllm_model.model.score(query, documents)
+    outputs = model.score(query, documents)
 
     return [output.outputs.score for output in outputs]
 
 
-def hf_reranker(
-    hf_runner: type[HfRunner],
-    model_name: str,
-    dtype: str,
-    query_strs: list[str],
-    document_strs: list[str],
-    query_type: str = "text",
-    doc_type: str = "text",
-):
+def hf_reranker(model_name,
+                query,
+                documents,
+                query_type="text",
+                doc_type="text"):
+
     checkpoint_to_hf_mapper = {
         "visual.": "model.visual.",
         "model.": "model.language_model.",
     }
 
-    data_pairs = [[query_strs[0], d] for d in document_strs]
+    model = AutoModel.from_pretrained(
+        model_name,
+        torch_dtype="auto",
+        trust_remote_code=True,
+        key_mapping=checkpoint_to_hf_mapper).to("cuda").eval()
 
-    with hf_runner(
-            model_name,
-            dtype=dtype,
-            trust_remote_code=True,
-            auto_cls=AutoModel,
-            model_kwargs={"key_mapping": checkpoint_to_hf_mapper},
-    ) as hf_model:
-        return hf_model.model.compute_score(data_pairs,
-                                            max_length=2048,
-                                            query_type=query_type,
-                                            doc_type=doc_type)
+    data_pairs = [[query[0], d] for d in documents]
+
+    scores = model.compute_score(data_pairs,
+                                 max_length=2048,
+                                 query_type=query_type,
+                                 doc_type=doc_type)
+    return scores
 
 
 # Visual Documents Reranking
 @pytest.mark.parametrize("model_name", [model_name])
-@pytest.mark.parametrize("dtype", ["half"])
-def test_model_text_image(hf_runner, vllm_runner, model_name, dtype):
+def test_model_text_image(model_name):
+
     query = ["slm markdown"]
     documents = [
         "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/handelsblatt-preview.png",
         "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png",
     ]
 
-    hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
-                             "text", "image")
-    vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
-                                 documents, "text", "image")
+    hf_outputs = hf_reranker(model_name, query, documents, "text", "image")
+    vllm_outputs = vllm_reranker(model_name, query, documents, "text", "image")
 
     assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
     assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -111,8 +88,8 @@ def test_model_text_image(hf_runner, vllm_runner, model_name, dtype):
 
 # Textual Documents Reranking
 @pytest.mark.parametrize("model_name", [model_name])
-@pytest.mark.parametrize("dtype", ["half"])
-def test_model_text_text(hf_runner, vllm_runner, model_name, dtype):
+def test_model_text_text(model_name):
+
     query = ["slm markdown"]
     documents = [
         """We present ReaderLM-v2, a compact 1.5 billion parameter language model designed for efficient 
@@ -127,10 +104,9 @@ def test_model_text_text(hf_runner, vllm_runner, model_name, dtype):
         lower computational requirements.""",  # noqa: E501
         "数据提取么？为什么不用正则啊，你用正则不就全解决了么？",
     ]
-    hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
-                             "text", "text")
-    vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
-                                 documents, "text", "text")
+
+    hf_outputs = hf_reranker(model_name, query, documents, "text", "text")
+    vllm_outputs = vllm_reranker(model_name, query, documents, "text", "text")
 
     assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
     assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -138,8 +114,8 @@ def test_model_text_text(hf_runner, vllm_runner, model_name, dtype):
 
 # Image Querying for Textual Documents
 @pytest.mark.parametrize("model_name", [model_name])
-@pytest.mark.parametrize("dtype", ["half"])
-def test_model_image_text(hf_runner, vllm_runner, model_name, dtype):
+def test_model_image_text(model_name):
+
     query = [
         "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png"
     ]
@@ -157,10 +133,8 @@ def test_model_image_text(hf_runner, vllm_runner, model_name, dtype):
         "数据提取么？为什么不用正则啊，你用正则不就全解决了么？",
     ]
 
-    hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
-                             "image", "text")
-    vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
-                                 documents, "image", "text")
+    hf_outputs = hf_reranker(model_name, query, documents, "image", "text")
+    vllm_outputs = vllm_reranker(model_name, query, documents, "image", "text")
 
     assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
     assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)
@@ -168,8 +142,8 @@ def test_model_image_text(hf_runner, vllm_runner, model_name, dtype):
 
 # Image Querying for Image Documents
 @pytest.mark.parametrize("model_name", [model_name])
-@pytest.mark.parametrize("dtype", ["half"])
-def test_model_image_image(hf_runner, vllm_runner, model_name, dtype):
+def test_model_image_image(model_name):
+
     query = [
         "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png"
     ]
@@ -178,10 +152,9 @@ def test_model_image_image(hf_runner, vllm_runner, model_name, dtype):
         "https://raw.githubusercontent.com/jina-ai/multimodal-reranker-test/main/paper-11.png",
     ]
 
-    hf_outputs = hf_reranker(hf_runner, model_name, dtype, query, documents,
-                             "image", "image")
-    vllm_outputs = vllm_reranker(vllm_runner, model_name, dtype, query,
-                                 documents, "image", "image")
+    hf_outputs = hf_reranker(model_name, query, documents, "image", "image")
+    vllm_outputs = vllm_reranker(model_name, query, documents, "image",
+                                 "image")
 
     assert hf_outputs[0] == pytest.approx(vllm_outputs[0], rel=0.02)
     assert hf_outputs[1] == pytest.approx(vllm_outputs[1], rel=0.02)

tests/models/multimodal/processing/test_common.py

@@ -159,7 +159,6 @@ def _test_processing_correctness(
 _ADD_SPECIAL_TOKENS_OVERRIDES = {
     "mllama": False,
     "ovis": False,
-    "paligemma": False,
     "ultravox": False,
     "whisper": False,
 }

tests/models/registry.py

@@ -141,8 +141,6 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                          trust_remote_code=True),
     "BaichuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan2-7B-chat",
                                          trust_remote_code=True),
-    "BailingMoeForCausalLM": _HfExamplesInfo("inclusionAI/Ling-lite-1.5",
-                                         trust_remote_code=True),
     "BambaForCausalLM": _HfExamplesInfo("ibm-ai-platform/Bamba-9B",
                                         extras={"tiny": "hmellor/tiny-random-BambaForCausalLM"}),  # noqa: E501
     "BloomForCausalLM": _HfExamplesInfo("bigscience/bloom-560m",
@@ -250,10 +248,6 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Phi3SmallForCausalLM": _HfExamplesInfo("microsoft/Phi-3-small-8k-instruct",
                                             trust_remote_code=True,
                                             v0_only=True),
-    "Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
-                                        trust_remote_code=True,
-                                        v0_only=True,
-                                        max_model_len=10240),
     "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                          trust_remote_code=True),
     "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
@@ -332,7 +326,6 @@ _CROSS_ENCODER_EXAMPLE_MODELS = {
                                                       hf_overrides={"architectures": ["GemmaForSequenceClassification"], # noqa: E501
                                                                     "classifier_from_token": ["Yes"], # noqa: E501
                                                                     "method": "no_post_processing"}), # noqa: E501
-    "LlamaForSequenceClassification": _HfExamplesInfo("Skywork/Skywork-Reward-V2-Llama-3.2-1B"), # noqa: E501
     "ModernBertForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-reranker-modernbert-base", v0_only=True), # noqa: E501
     "RobertaForSequenceClassification": _HfExamplesInfo("cross-encoder/quora-roberta-base", v0_only=True),  # noqa: E501
     "XLMRobertaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-m3", v0_only=True),  # noqa: E501
@@ -440,7 +433,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
                                                          tokenizer="Isotr0py/Florence-2-tokenizer",  # noqa: E501
                                                          trust_remote_code=True),  # noqa: E501
     "MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"),  # noqa: E501
-    "VoxtralForConditionalGeneration": _HfExamplesInfo("mistralai/Voxtral-Mini-3B-2507", tokenizer_mode="mistral"),  # noqa: E501
     "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
 
     # [Cross-encoder]
@@ -465,11 +457,6 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
                                             trust_remote_code=True,
                                             speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
                                             tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
-    "EagleLlama4ForCausalLM": _HfExamplesInfo(
-        "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
-        trust_remote_code=True,
-        speculative_model="morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
-        tokenizer="meta-llama/Llama-4-Scout-17B-16E-Instruct"),  # noqa: E501
     "EagleMiniCPMForCausalLM": _HfExamplesInfo("openbmb/MiniCPM-1B-sft-bf16",
                                             trust_remote_code=True,
                                             is_available_online=False,
@@ -519,4 +506,4 @@ class HfExampleModels:
         raise ValueError(f"No example model defined for {model_id}")
 
 
-HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
+HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)

tests/models/test_initialization.py

@@ -31,16 +31,10 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
     model_info.check_transformers_version(on_fail="skip")
 
     # FIXME: Possible memory leak in the previous tests?
-    if model_arch in ("Glm4vForConditionalGeneration",
-                      "GraniteSpeechForConditionalGeneration",
+    if model_arch in ("GraniteSpeechForConditionalGeneration",
                       "KimiVLForConditionalGeneration"):
         pytest.skip("Avoid OOM")
 
-    if model_arch in ("Llama4ForCausalLM", "EagleLlama4ForCausalLM"):
-        from vllm.model_executor.models.llama4 import Llama4ForCausalLM
-        from vllm.model_executor.models.registry import ModelRegistry
-        ModelRegistry.register_model("Llama4ForCausalLM", Llama4ForCausalLM)
-
     # Avoid OOM and reduce initialization time by only using 1 layer
     def hf_overrides(hf_config: PretrainedConfig) -> PretrainedConfig:
         hf_config.update(model_info.hf_overrides)
@@ -48,18 +42,13 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
         text_config = hf_config.get_text_config()
 
         # Ensure at least 2 expert per group
-        # Since `grouped_topk` assumes top-2
+        # Since `grouped_topk` assums top-2
         n_group = getattr(text_config, 'n_group', None)
         num_experts = n_group * 2 if n_group is not None else 2
 
-        # we use three layers for Gemma-3n to check
-        # both normal layer and kv_shared_layer
-        num_hidden_layers = (3 if model_arch
-                             == "Gemma3nForConditionalGeneration" else 1)
-
         text_config.update({
             "num_layers": 1,
-            "num_hidden_layers": num_hidden_layers,
+            "num_hidden_layers": 1,
             "num_experts": num_experts,
             "num_experts_per_tok": 2,
             "num_local_experts": num_experts,
@@ -67,8 +56,6 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
             "first_k_dense_replace": 0,
             # To avoid OOM on DeepSeek-V3
             "n_routed_experts": num_experts,
-            # For Gemma-3n
-            "num_kv_shared_layers": 1,
         })
 
         if hasattr(hf_config, "vision_config"):
@@ -108,9 +95,6 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
                        _initialize_kv_caches_v1), monkeypatch.context() as m):
         if model_info.v0_only:
             m.setenv("VLLM_USE_V1", "0")
-        if model_arch == "Phi4FlashForCausalLM":
-            # Phi4FlashForCausalLM only supports DIFFERENTIAL_FLASH_ATTN backend
-            m.setenv("VLLM_ATTENTION_BACKEND", "DIFFERENTIAL_FLASH_ATTN")
         LLM(
             model_info.default,
             tokenizer=model_info.tokenizer,

tests/multimodal/test_utils.py

@@ -39,7 +39,7 @@ TEST_IMAGE_URLS = [
 
 TEST_VIDEO_URLS = [
     "https://www.bogotobogo.com/python/OpenCV_Python/images/mean_shift_tracking/slow_traffic_small.mp4",
-    "https://github.com/opencv/opencv/raw/refs/tags/4.12.0/samples/data/vtest.avi",
+    "https://filesamples.com/samples/video/avi/sample_640x360.avi",
 ]
 
 

tests/test_config.py

@@ -7,7 +7,7 @@ import pytest
 
 from vllm.compilation.backends import VllmBackend
 from vllm.config import (LoadConfig, ModelConfig, PoolerConfig, VllmConfig,
-                         get_field, update_config)
+                         get_field)
 from vllm.model_executor.layers.pooler import PoolingType
 from vllm.platforms import current_platform
 
@@ -46,34 +46,6 @@ def test_get_field():
     assert c.default_factory is MISSING
 
 
-@dataclass
-class _TestNestedConfig:
-    a: _TestConfigFields = field(
-        default_factory=lambda: _TestConfigFields(a=0))
-
-
-def test_update_config():
-    # Simple update
-    config1 = _TestConfigFields(a=0)
-    new_config1 = update_config(config1, {"a": 42})
-    assert new_config1.a == 42
-    # Nonexistent field
-    with pytest.raises(AssertionError):
-        new_config1 = update_config(config1, {"nonexistent": 1})
-    # Nested update with dataclass
-    config2 = _TestNestedConfig()
-    new_inner_config = _TestConfigFields(a=1, c="new_value")
-    new_config2 = update_config(config2, {"a": new_inner_config})
-    assert new_config2.a == new_inner_config
-    # Nested update with dict
-    config3 = _TestNestedConfig()
-    new_config3 = update_config(config3, {"a": {"c": "new_value"}})
-    assert new_config3.a.c == "new_value"
-    # Nested update with invalid type
-    with pytest.raises(AssertionError):
-        new_config3 = update_config(config3, {"a": "new_value"})
-
-
 @pytest.mark.parametrize(
     ("model_id", "expected_runner_type", "expected_task"),
     [
@@ -82,7 +54,7 @@ def test_update_config():
         ("jason9693/Qwen2.5-1.5B-apeach", "pooling", "classify"),
         ("cross-encoder/ms-marco-MiniLM-L-6-v2", "pooling", "classify"),
         ("Qwen/Qwen2.5-Math-RM-72B", "pooling", "reward"),
-        ("openai/whisper-small", "generate", "transcription"),
+        ("openai/whisper-small", "transcription", "transcription"),
     ],
 )
 def test_auto_task(model_id, expected_runner_type, expected_task):
@@ -97,11 +69,7 @@ def test_auto_task(model_id, expected_runner_type, expected_task):
     )
 
     assert config.runner_type == expected_runner_type
-
-    if config.runner_type == "pooling":
-        assert config.task == expected_task
-    else:
-        assert expected_task in config.supported_tasks
+    assert config.task == expected_task
 
 
 @pytest.mark.parametrize(
@@ -130,50 +98,11 @@ def test_score_task(model_id, expected_runner_type, expected_task):
     assert config.task == expected_task
 
 
-@pytest.mark.parametrize(("model_id", "expected_runner_type", "expected_task"),
-                         [
-                             ("Qwen/Qwen2.5-1.5B-Instruct", "draft", "auto"),
-                         ])
-def test_draft_task(model_id, expected_runner_type, expected_task):
-    config = ModelConfig(
-        model_id,
-        runner="draft",
-        tokenizer=model_id,
-        seed=0,
-        dtype="float16",
-    )
-
-    assert config.runner_type == expected_runner_type
-    assert config.task == expected_task
-
-
-@pytest.mark.parametrize(
-    ("model_id", "expected_runner_type", "expected_task"),
-    [
-        ("openai/whisper-small", "generate", "transcription"),
-    ],
-)
-def test_transcription_task(model_id, expected_runner_type, expected_task):
-    config = ModelConfig(
-        model_id,
-        task="transcription",
-        tokenizer=model_id,
-        tokenizer_mode="auto",
-        trust_remote_code=False,
-        seed=0,
-        dtype="float16",
-    )
-
-    assert config.runner_type == expected_runner_type
-    assert config.task == expected_task
-
-
 @pytest.mark.parametrize(("model_id", "bad_task"), [
     ("Qwen/Qwen2.5-Math-RM-72B", "generate"),
-    ("Qwen/Qwen3-0.6B", "transcription"),
 ])
 def test_incorrect_task(model_id, bad_task):
-    with pytest.raises(ValueError, match=r"does not support task=.*"):
+    with pytest.raises(ValueError, match=r"does not support the .* task"):
         ModelConfig(
             model_id,
             task=bad_task,

tests/test_utils.py

@@ -458,31 +458,6 @@ def test_bind_kv_cache():
     assert ctx['layers.2.self_attn'].kv_cache[0] is kv_cache[2]
     assert ctx['layers.3.self_attn'].kv_cache[0] is kv_cache[3]
 
-def test_bind_kv_cache_kv_sharing():
-    from vllm.attention import Attention
-
-    ctx = {
-        'layers.0.self_attn': Attention(32, 128, 0.1),
-        'layers.1.self_attn': Attention(32, 128, 0.1),
-        'layers.2.self_attn': Attention(32, 128, 0.1),
-        'layers.3.self_attn': Attention(32, 128, 0.1),
-    }
-    kv_cache = [
-        torch.zeros((1, )),
-        torch.zeros((1, )),
-        torch.zeros((1, )),
-        torch.zeros((1, )),
-    ]
-    shared_kv_cache_layers = {
-        'layers.2.self_attn': 'layers.1.self_attn',
-        'layers.3.self_attn': 'layers.0.self_attn'
-    }
-    bind_kv_cache(ctx, [kv_cache], shared_kv_cache_layers)
-    assert ctx['layers.0.self_attn'].kv_cache[0] is kv_cache[0]
-    assert ctx['layers.1.self_attn'].kv_cache[0] is kv_cache[1]
-    assert ctx['layers.2.self_attn'].kv_cache[0] is kv_cache[1]
-    assert ctx['layers.3.self_attn'].kv_cache[0] is kv_cache[0]
-
 def test_bind_kv_cache_non_attention():
     from vllm.attention import Attention
 

tests/tool_use/test_kimi_k2_tool_parser.py

@@ -10,6 +10,8 @@ from vllm.entrypoints.openai.protocol import FunctionCall, ToolCall
 from vllm.entrypoints.openai.tool_parsers import KimiK2ToolParser
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
+pytest.skip("skip kimi_k2 parser test", allow_module_level=True)
+
 # Use a common model that is likely to be available
 MODEL = "moonshotai/Kimi-K2-Instruct"
 

tests/tpu/test_quantization_accuracy.py

@@ -14,7 +14,7 @@ RTOL = 0.03
 @dataclass
 class GSM8KAccuracyTestConfig:
     model_name: str
-    expected_value: float
+    excepted_value: float
 
     def get_model_args(self) -> str:
         return (f"pretrained={self.model_name},"
@@ -25,13 +25,13 @@ class GSM8KAccuracyTestConfig:
 ACCURACY_CONFIGS = [
     GSM8KAccuracyTestConfig(
         model_name="neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        expected_value=0.76),  # no bias
+        excepted_value=0.76),  # no bias
     # NOTE(rob): We cannot re-initialize vLLM in the same process for TPU,
     # so only one of these tests can run in a single call to pytest. As
     # a follow up, move this into the LM-EVAL section of the CI.
     # GSM8KAccuracyTestConfig(
     #     model_name="neuralmagic/Qwen2-7B-Instruct-quantized.w8a8",
-    #     expected_value=0.66),  # bias in QKV layers
+    #     excepted_value=0.66),  # bias in QKV layers
 ]
 
 
@@ -45,7 +45,7 @@ def test_gsm8k_correctness(config: GSM8KAccuracyTestConfig):
         batch_size="auto",
     )
 
-    EXPECTED_VALUE = config.expected_value
+    EXPECTED_VALUE = config.excepted_value
     measured_value = results["results"][TASK][FILTER]
     assert (measured_value - RTOL < EXPECTED_VALUE
             and measured_value + RTOL > EXPECTED_VALUE

tests/v1/core/test_async_scheduler.py

@@ -1,228 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from collections import deque
-
-import pytest
-
-from vllm.v1.core.sched.output import SchedulerOutput
-from vllm.v1.outputs import ModelRunnerOutput
-from vllm.v1.request import RequestStatus
-
-from .utils import create_requests, create_scheduler
-
-
-def _make_model_runner_output(
-    scheduler_output: SchedulerOutput, ) -> ModelRunnerOutput:
-    req_ids = list(scheduler_output.num_scheduled_tokens.keys())
-    return ModelRunnerOutput(
-        req_ids=req_ids,
-        req_id_to_index={
-            req_id: i
-            for i, req_id in enumerate(req_ids)
-        },
-        sampled_token_ids=[[i] for i in range(len(req_ids))],
-        spec_token_ids=None,
-        logprobs=None,
-        prompt_logprobs_dict={},
-        pooler_output=[],
-    )
-
-
-@pytest.mark.parametrize("max_tokens", [1, 2, 3, 5])
-def test_stop_by_max_tokens(max_tokens: int):
-    scheduler = create_scheduler(async_scheduling=True)
-    requests = create_requests(num_requests=2, max_tokens=max_tokens)
-    req0, req1 = requests
-
-    sched_outputs: deque[SchedulerOutput] = deque()
-    scheduler.add_request(req0)
-    sched_outputs.append(scheduler.schedule())
-
-    scheduler.add_request(req1)
-    sched_outputs.append(scheduler.schedule())
-
-    while sched_outputs:
-        sched_output = sched_outputs.popleft()
-        model_runner_output = _make_model_runner_output(sched_output)
-        scheduler.update_from_output(sched_output, model_runner_output)
-
-        sched_output = scheduler.schedule()
-        if sched_output.num_scheduled_tokens:
-            sched_outputs.append(sched_output)
-
-    assert scheduler.get_num_unfinished_requests() == 0
-    assert req0.num_output_tokens == max_tokens
-    assert req1.num_output_tokens == max_tokens
-
-
-def test_abort():
-    scheduler = create_scheduler(async_scheduling=True)
-    requests = create_requests(num_requests=10, max_tokens=20)
-
-    for req in requests:
-        scheduler.add_request(req)
-
-    sched_outputs: deque[SchedulerOutput] = deque()
-    sched_outputs.append(scheduler.schedule())
-    sched_outputs.append(scheduler.schedule())
-
-    abort_order = [0, 8, 3, 1, 6, 4, 2, 5, 7, 9]
-    abort_order_copy = abort_order.copy()
-
-    def abort_request():
-        if not abort_order:
-            return
-        req = requests[abort_order.pop(0)]
-        scheduler.finish_requests(req.request_id,
-                                  RequestStatus.FINISHED_ABORTED)
-
-    while sched_outputs:
-        # Abort a scheduled request.
-        abort_request()
-        sched_output = sched_outputs.popleft()
-        model_runner_output = _make_model_runner_output(sched_output)
-        scheduler.update_from_output(sched_output, model_runner_output)
-
-        sched_output = scheduler.schedule()
-        if sched_output.num_scheduled_tokens:
-            sched_outputs.append(sched_output)
-
-    for i, req in enumerate(requests):
-        assert req.status == RequestStatus.FINISHED_ABORTED
-        assert req.num_output_tokens == abort_order_copy.index(i)
-
-
-def test_preempt():
-    scheduler = create_scheduler(async_scheduling=True)
-    requests = create_requests(num_requests=10, max_tokens=20)
-
-    for req in requests:
-        scheduler.add_request(req)
-
-    sched_outputs: deque[SchedulerOutput] = deque()
-    sched_outputs.append(scheduler.schedule())
-    sched_outputs.append(scheduler.schedule())
-
-    abort_order = [0, 8, 3, 1, 6, 4, 2, 5, 7, 9]
-    abort_order_copy = abort_order.copy()
-
-    def abort_request():
-        if not abort_order:
-            return
-        req = requests[abort_order.pop(0)]
-        scheduler.finish_requests(req.request_id,
-                                  RequestStatus.FINISHED_ABORTED)
-
-    while sched_outputs:
-        # Abort a scheduled request.
-        abort_request()
-        sched_output = sched_outputs.popleft()
-        model_runner_output = _make_model_runner_output(sched_output)
-        scheduler.update_from_output(sched_output, model_runner_output)
-
-        sched_output = scheduler.schedule()
-        if sched_output.num_scheduled_tokens:
-            sched_outputs.append(sched_output)
-
-    for i, req in enumerate(requests):
-        assert req.status == RequestStatus.FINISHED_ABORTED
-        assert req.num_output_tokens == abort_order_copy.index(i)
-
-
-def test_prefix_caching_for_prefill_dedup():
-    CHUNK_SIZE = 1000
-    BLOCK_SIZE = 16
-    num_prompt_tokens = 100
-    scheduler = create_scheduler(async_scheduling=True,
-                                 max_num_batched_tokens=CHUNK_SIZE,
-                                 enable_prefix_caching=True,
-                                 block_size=BLOCK_SIZE)
-    requests = create_requests(num_requests=5,
-                               num_tokens=num_prompt_tokens,
-                               max_tokens=3,
-                               same_prompt=True)
-    requests_copy = requests.copy()
-
-    # Two requests with the same prompt.
-    req0 = requests.pop(0)
-    req1 = requests.pop(0)
-    scheduler.add_request(req0)
-    scheduler.add_request(req1)
-
-    sched_outputs: deque[SchedulerOutput] = deque()
-    sched_output = scheduler.schedule()
-    sched_outputs.append(sched_output)
-    # Make sure prefix caching de-duplicates the prompts in the same step,
-    # so all the blocks except the last are shared between the two requests.
-    assert len(sched_output.num_scheduled_tokens) == 2
-    num_blocks = num_prompt_tokens // BLOCK_SIZE
-    assert req0.num_cached_tokens == 0
-    assert req1.num_cached_tokens >= num_blocks * BLOCK_SIZE
-
-    sched_outputs.append(scheduler.schedule())
-    while sched_outputs:
-        if requests:
-            scheduler.add_request(requests.pop(0))
-        sched_output = sched_outputs.popleft()
-        model_runner_output = _make_model_runner_output(sched_output)
-        scheduler.update_from_output(sched_output, model_runner_output)
-        sched_output = scheduler.schedule()
-        if sched_output.num_scheduled_tokens:
-            sched_outputs.append(sched_output)
-
-    # Other requests scheduled after the two requests should also get
-    # prefix cache hit.
-    assert scheduler.get_num_unfinished_requests() == 0
-    for req in requests_copy[1:]:
-        assert req.num_cached_tokens >= num_blocks * BLOCK_SIZE
-
-
-def test_prefix_caching_for_multi_turn():
-    CHUNK_SIZE = 1000
-    BLOCK_SIZE = 16
-    num_prompt_tokens = 100
-    num_output_tokens = 200
-    scheduler = create_scheduler(async_scheduling=True,
-                                 max_num_batched_tokens=CHUNK_SIZE,
-                                 enable_prefix_caching=True,
-                                 block_size=BLOCK_SIZE)
-    requests = create_requests(num_requests=5,
-                               num_tokens=num_prompt_tokens,
-                               max_tokens=num_output_tokens)
-
-    for req in requests:
-        scheduler.add_request(req)
-    sched_outputs: deque[SchedulerOutput] = deque()
-    sched_outputs.append(scheduler.schedule())
-    sched_outputs.append(scheduler.schedule())
-
-    # Process the requests.
-    while sched_outputs:
-        sched_output = sched_outputs.popleft()
-        model_runner_output = _make_model_runner_output(sched_output)
-        scheduler.update_from_output(sched_output, model_runner_output)
-        sched_output = scheduler.schedule()
-        if sched_output.num_scheduled_tokens:
-            sched_outputs.append(sched_output)
-    assert scheduler.get_num_unfinished_requests() == 0
-
-    # Create next-turn requests whose prompts are the full output of the
-    # previous turn.
-    next_turn_requests = create_requests(
-        num_requests=5,
-        num_tokens=num_prompt_tokens + num_output_tokens,
-        max_tokens=num_output_tokens,
-    )
-    for i, req in enumerate(next_turn_requests):
-        req.prompt_token_ids = (requests[i].prompt_token_ids +
-                                list(requests[i].output_token_ids))
-    # Schedule the next-turn requests.
-    for req in next_turn_requests:
-        scheduler.add_request(req)
-    sched_outputs.append(scheduler.schedule())
-
-    # Make sure the next-turn requests get prefix cache hit by the previous
-    # requests.
-    for req in next_turn_requests:
-        assert (req.num_cached_tokens == req.num_prompt_tokens // BLOCK_SIZE *
-                BLOCK_SIZE)

tests/v1/core/test_kv_cache_utils.py

@@ -8,7 +8,7 @@ import torch
 from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
 from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
 from vllm.sampling_params import SamplingParams
-from vllm.utils import GiB_bytes, sha256, sha256_cbor_64bit
+from vllm.utils import GiB_bytes, sha256
 from vllm.v1.core.kv_cache_manager import KVCacheManager
 # disable yapf here as it formats differently than isort such that both fail
 # yapf: disable
@@ -16,8 +16,7 @@ from vllm.v1.core.kv_cache_utils import (
     FreeKVCacheBlockQueue, KVCacheBlock, PrefixCachingMetrics,
     estimate_max_model_len, generate_block_hash_extra_keys,
     get_kv_cache_config, get_max_concurrency_for_kv_cache_config,
-    hash_block_tokens, hash_request_tokens, init_none_hash,
-    unify_kv_cache_configs)
+    hash_block_tokens, hash_request_tokens, unify_kv_cache_configs)
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheGroupSpec, KVCacheTensor,
                                         SlidingWindowSpec)
@@ -79,27 +78,24 @@ def new_sliding_window_spec(block_size=16,
                              sliding_window=sliding_window)
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
-def test_none_hash(monkeypatch, hash_fn):
+def test_none_hash(monkeypatch):
     import vllm.v1.core.kv_cache_utils
 
     # case 1: PYTHONHASHSEED is not set, use random
     with monkeypatch.context() as m:
         m.delenv('PYTHONHASHSEED', raising=False)
         reloaded_kv_cache_utils = importlib.reload(vllm.v1.core.kv_cache_utils)
-        reloaded_kv_cache_utils.init_none_hash(hash_fn)
         assert reloaded_kv_cache_utils.NONE_HASH is not None
         assert isinstance(reloaded_kv_cache_utils.NONE_HASH, int)
         assert reloaded_kv_cache_utils.NONE_HASH != 0
 
-    # case 2: PYTHONHASHSEED is set, use the seed and hash_fn
+    # case 2: PYTHONHASHSEED is set, use the seed
     with monkeypatch.context() as m:
         m.setenv('PYTHONHASHSEED', 'python hash seed')
         reloaded_kv_cache_utils = importlib.reload(vllm.v1.core.kv_cache_utils)
-        reloaded_kv_cache_utils.init_none_hash(hash_fn)
         assert reloaded_kv_cache_utils.NONE_HASH is not None
         assert isinstance(reloaded_kv_cache_utils.NONE_HASH, int)
-        assert hash_fn('python hash seed') == reloaded_kv_cache_utils.NONE_HASH
+        assert sha256('python hash seed') == reloaded_kv_cache_utils.NONE_HASH
 
 
 def test_kv_cache_block():
@@ -291,10 +287,9 @@ def test_generate_block_hash_extra_keys_cache_salt():
     assert next_mm_idx == 1
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
+@pytest.mark.parametrize("hash_fn", [sha256, hash])
 def test_hash_block_tokens(hash_fn):
     import vllm.v1.core.kv_cache_utils
-    init_none_hash(hash_fn)
     parent_block_hash = 123
     curr_block_token_ids = (1, 2, 3)
     extra_keys = ("key1", "key2")
@@ -308,10 +303,9 @@ def test_hash_block_tokens(hash_fn):
     assert block_hash.extra_keys == extra_keys
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
+@pytest.mark.parametrize("hash_fn", [sha256, hash])
 def test_hash_request_tokens(hash_fn):
     import vllm.v1.core.kv_cache_utils
-    init_none_hash(hash_fn)
     request = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(6)],
@@ -338,10 +332,8 @@ def test_hash_request_tokens(hash_fn):
     assert block_hashes[1].extra_keys == ("hash2", )
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
+@pytest.mark.parametrize("hash_fn", [sha256, hash])
 def test_hash_tokens_different_mm_input(hash_fn):
-    init_none_hash(hash_fn)
-
     request1 = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(6)],
@@ -367,10 +359,8 @@ def test_hash_tokens_different_mm_input(hash_fn):
     assert block_hashes1[1] != block_hashes2[1]
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
+@pytest.mark.parametrize("hash_fn", [sha256, hash])
 def test_hash_request_tokens_no_mm_inputs(hash_fn):
-    init_none_hash(hash_fn)
-
     request = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(6)],
@@ -926,4 +916,4 @@ def test_get_kv_cache_config():
         ],
         kv_cache_groups=[
             KVCacheGroupSpec(["layer_1", "layer_2"], new_kv_cache_spec())
-        ])
+        ])

tests/v1/core/test_prefix_caching.py

@@ -11,12 +11,11 @@ import torch
 from vllm.distributed.kv_events import AllBlocksCleared, BlockRemoved
 from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
 from vllm.sampling_params import SamplingParams
-from vllm.utils import sha256, sha256_cbor_64bit
+from vllm.utils import sha256
 from vllm.v1.core.block_pool import BlockPool
 from vllm.v1.core.kv_cache_manager import KVCacheManager, Request
 from vllm.v1.core.kv_cache_utils import (BlockHash, BlockHashWithGroupId,
-                                         KVCacheBlock, hash_block_tokens,
-                                         init_none_hash)
+                                         KVCacheBlock, hash_block_tokens)
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
                                         KVCacheGroupSpec, SlidingWindowSpec)
 
@@ -92,7 +91,7 @@ def make_kv_cache_config_hybrid_model(block_size: int,
     )
 
 
-@pytest.mark.parametrize("hash_algo", ["sha256", "sha256_cbor_64bit", "hash"])
+@pytest.mark.parametrize("hash_algo", ["sha256", "hash"])
 def test_prefill(hash_algo):
     manager = KVCacheManager(
         make_kv_cache_config(16, 11),
@@ -102,8 +101,7 @@ def test_prefill(hash_algo):
     )
 
     # choose the hash function according to the parameter
-    hash_fn = (sha256_cbor_64bit if hash_algo == "sha256_cbor_64bit" else
-               sha256 if hash_algo == "sha256" else hash)
+    hash_fn = sha256 if hash_algo == "sha256" else hash
 
     # Complete 3 blocks (48 tokens)
     common_token_ids = [i for i in range(3) for _ in range(16)]
@@ -698,14 +696,12 @@ def test_basic_prefix_caching_disabled():
     assert not blocks
 
 
-@pytest.mark.parametrize("hash_fn", [sha256, sha256_cbor_64bit, hash])
+@pytest.mark.parametrize("hash_fn", [sha256, hash])
 def test_cache_blocks(hash_fn):
     """
     This is a unit test that tests the correctness of the _cache_full_blocks
     function of KVCacheManager.
     """
-    init_none_hash(hash_fn)
-
     block_size = 4
     block_pool = BlockPool(
         num_gpu_blocks=5,

tests/v1/core/test_scheduler.py

@@ -19,7 +19,133 @@ from vllm.v1.request import Request, RequestStatus
 from vllm.v1.structured_output import StructuredOutputManager
 from vllm.v1.structured_output.request import StructuredOutputRequest
 
-from .utils import EOS_TOKEN_ID, create_requests, create_scheduler
+EOS_TOKEN_ID = 50256
+
+
+def create_scheduler(
+    model: str = "facebook/opt-125m",
+    max_num_seqs: int = 16,
+    max_num_batched_tokens: int = 8192,
+    enable_prefix_caching: Optional[bool] = None,
+    long_prefill_token_threshold: int = 0,
+    disable_chunked_mm_input: bool = False,
+    use_kv_connector: bool = False,
+    num_blocks: int = 10000,
+    block_size: int = 16,
+    max_model_len: Optional[int] = None,
+    num_speculative_tokens: Optional[int] = None,
+    skip_tokenizer_init: bool = False,
+) -> Scheduler:
+    '''Create scheduler under test.
+
+    Args:
+      model: model under test
+      max_num_seqs: max sequences to schedule
+      max_num_batch_tokens: max num tokens to batch
+      enable_prefix_caching: optionally force APC config
+                             (True/False) or use default
+                             (None)
+
+    Returns:
+      {class}`Scheduler` instance
+    '''
+    if max_model_len is None:
+        max_model_len = max_num_batched_tokens
+    scheduler_config = SchedulerConfig(
+        max_num_seqs=max_num_seqs,
+        max_num_batched_tokens=max_num_batched_tokens,
+        max_model_len=max_model_len,
+        long_prefill_token_threshold=long_prefill_token_threshold,
+        disable_chunked_mm_input=disable_chunked_mm_input,
+        enable_chunked_prefill=True,
+    )
+    model_config = ModelConfig(
+        model=model,
+        task="auto",
+        tokenizer=model,
+        tokenizer_mode="auto",
+        trust_remote_code=True,
+        dtype="float16",
+        seed=42,
+        skip_tokenizer_init=skip_tokenizer_init,
+    )
+    # Cache config, optionally force APC
+    kwargs_cache = ({} if enable_prefix_caching is None else {
+        'enable_prefix_caching': enable_prefix_caching
+    })
+    cache_config = CacheConfig(
+        block_size=block_size,
+        gpu_memory_utilization=0.9,
+        swap_space=0,
+        cache_dtype="auto",
+        **kwargs_cache,
+    )
+    kv_transfer_config = KVTransferConfig(
+        kv_connector="SharedStorageConnector",
+        kv_role="kv_both",
+        kv_connector_extra_config={"shared_storage_path": "local_storage"},
+    ) if use_kv_connector else None
+
+    speculative_config: Optional[SpeculativeConfig] = None
+    if num_speculative_tokens is not None:
+        speculative_config = SpeculativeConfig(
+            model="ngram", num_speculative_tokens=num_speculative_tokens)
+
+    vllm_config = VllmConfig(
+        scheduler_config=scheduler_config,
+        model_config=model_config,
+        cache_config=cache_config,
+        kv_transfer_config=kv_transfer_config,
+        speculative_config=speculative_config,
+    )
+    kv_cache_config = KVCacheConfig(
+        num_blocks=num_blocks,  # A large number of blocks to hold all requests
+        kv_cache_tensors=[],
+        kv_cache_groups=[
+            KVCacheGroupSpec(['layer'],
+                             FullAttentionSpec(block_size, 1, 1, torch.float32,
+                                               False))
+        ],
+    )
+    cache_config.num_gpu_blocks = num_blocks
+    return Scheduler(
+        vllm_config=vllm_config,
+        kv_cache_config=kv_cache_config,
+        log_stats=True,
+        structured_output_manager=StructuredOutputManager(vllm_config),
+    )
+
+
+def create_requests(num_requests: int,
+                    num_tokens: int = 10,
+                    mm_positions: Optional[list[PlaceholderRange]] = None,
+                    max_tokens: int = 16,
+                    stop_token_ids: Optional[list[int]] = None,
+                    prompt_logprobs: Optional[int] = None):
+    sampling_params = SamplingParams(ignore_eos=False,
+                                     max_tokens=max_tokens,
+                                     stop_token_ids=stop_token_ids,
+                                     prompt_logprobs=prompt_logprobs)
+    requests = []
+    for i in range(num_requests):
+        if mm_positions is not None:
+            mm_position = mm_positions[i]
+            mm_inputs = [MultiModalKwargs({})] * len(mm_position)
+        else:
+            mm_position = None
+            mm_inputs = None
+        request = Request(
+            request_id=f"{i}",
+            prompt_token_ids=[i] * num_tokens,
+            sampling_params=sampling_params,
+            pooling_params=None,
+            multi_modal_inputs=mm_inputs,
+            multi_modal_placeholders=mm_position,
+            multi_modal_hashes=None,
+            eos_token_id=EOS_TOKEN_ID,
+        )
+        requests.append(request)
+    return requests
 
 
 def test_add_requests():
@@ -325,7 +451,6 @@ def test_stop_via_update_from_output():
         req.num_computed_tokens = req.num_tokens
         scheduler.requests[req.request_id] = req
         scheduler.running.append(req)
-        req.status = RequestStatus.RUNNING
 
     scheduler_output = SchedulerOutput(
         scheduled_new_reqs=[],
@@ -379,7 +504,6 @@ def test_stop_via_update_from_output():
         req.num_computed_tokens = req.num_tokens
         scheduler.requests[req.request_id] = req
         scheduler.running.append(req)
-        req.status = RequestStatus.RUNNING
 
     scheduler_output = SchedulerOutput(
         scheduled_new_reqs=[],
@@ -432,7 +556,6 @@ def test_stop_via_update_from_output():
         req.num_computed_tokens = req.num_tokens
         scheduler.requests[req.request_id] = req
         scheduler.running.append(req)
-        req.status = RequestStatus.RUNNING
 
     scheduler_output = SchedulerOutput(
         scheduled_new_reqs=[],
@@ -580,65 +703,6 @@ def test_schedule_concurrent_batches(enable_prefix_caching: Optional[bool],
     scheduler.update_from_output(scheduler_output1, model_runner_output)
 
 
-def test_preempt_during_execution():
-    # NOTE(woosuk): The actual number of available blocks is 10 instead of 11
-    # because block 0 is reserved as the null block.
-    scheduler = create_scheduler(max_num_batched_tokens=100,
-                                 block_size=16,
-                                 num_blocks=11,
-                                 enable_prefix_caching=False)
-    requests = create_requests(num_requests=2, num_tokens=80)
-
-    # Schedule the first request.
-    scheduler.add_request(requests[0])
-    scheduler_output0 = scheduler.schedule()
-    assert len(scheduler_output0.num_scheduled_tokens) == 1
-    assert len(scheduler_output0.scheduled_new_reqs[0].block_ids[0]) == 5
-
-    # Schedule the second request while the first request is still running.
-    # This scenario can occur in certain cases, when max_concurrent_batches > 1
-    # (e.g., when pipeline parallelism is used).
-    scheduler.add_request(requests[1])
-    scheduler_output1 = scheduler.schedule()
-    assert len(scheduler_output1.num_scheduled_tokens) == 1
-    assert len(scheduler_output1.scheduled_new_reqs[0].block_ids[0]) == 5
-
-    # Get the output of the first request.
-    model_runner_output0 = ModelRunnerOutput(
-        req_ids=[requests[0].request_id],
-        req_id_to_index={requests[0].request_id: 0},
-        sampled_token_ids=[[0]],
-        spec_token_ids=None,
-        logprobs=None,
-        prompt_logprobs_dict={},
-        pooler_output=[],
-    )
-    scheduler.update_from_output(scheduler_output0, model_runner_output0)
-
-    # Schedule the first request again. This will cause the preemption
-    # of the second request because the KV cache is full.
-    _ = scheduler.schedule()
-    assert len(scheduler.running) == 1
-    assert scheduler.running[0] == requests[0]
-    assert requests[1].status == RequestStatus.PREEMPTED
-
-    model_runner_output1 = ModelRunnerOutput(
-        req_ids=[requests[1].request_id],
-        req_id_to_index={requests[1].request_id: 0},
-        sampled_token_ids=[[42]],
-        spec_token_ids=None,
-        logprobs=None,
-        prompt_logprobs_dict={},
-        pooler_output=[],
-    )
-    scheduler.update_from_output(scheduler_output1, model_runner_output1)
-
-    # The second request (that is preempted) should be updated with the
-    # sampled token id.
-    assert len(requests[1].output_token_ids) == 1
-    assert requests[1].output_token_ids[0] == 42
-
-
 # Note - these test cases mirror some of those in test_rejection_sampler.py
 @pytest.mark.parametrize(
     "spec_tokens,output_tokens,expected",

tests/v1/core/utils.py

@@ -1,152 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from typing import Optional, Union
-
-import torch
-
-from vllm.config import (CacheConfig, KVTransferConfig, ModelConfig,
-                         SchedulerConfig, SpeculativeConfig, VllmConfig)
-from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
-from vllm.sampling_params import SamplingParams
-from vllm.v1.core.sched.async_scheduler import AsyncScheduler
-from vllm.v1.core.sched.scheduler import Scheduler
-from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
-                                        KVCacheGroupSpec)
-from vllm.v1.request import Request
-from vllm.v1.structured_output import StructuredOutputManager
-
-EOS_TOKEN_ID = 50256
-
-
-def create_scheduler(
-    model: str = "facebook/opt-125m",
-    max_num_seqs: int = 16,
-    max_num_batched_tokens: int = 8192,
-    enable_prefix_caching: Optional[bool] = None,
-    long_prefill_token_threshold: int = 0,
-    disable_chunked_mm_input: bool = False,
-    use_kv_connector: bool = False,
-    num_blocks: int = 10000,
-    block_size: int = 16,
-    max_model_len: Optional[int] = None,
-    num_speculative_tokens: Optional[int] = None,
-    skip_tokenizer_init: bool = False,
-    async_scheduling: bool = False,
-) -> Union[Scheduler, AsyncScheduler]:
-    '''Create scheduler under test.
-
-    Args:
-      model: model under test
-      max_num_seqs: max sequences to schedule
-      max_num_batch_tokens: max num tokens to batch
-      enable_prefix_caching: optionally force APC config
-                             (True/False) or use default
-                             (None)
-
-    Returns:
-      {class}`Scheduler` instance
-    '''
-    if max_model_len is None:
-        max_model_len = max_num_batched_tokens
-    scheduler_config = SchedulerConfig(
-        max_num_seqs=max_num_seqs,
-        max_num_batched_tokens=max_num_batched_tokens,
-        max_model_len=max_model_len,
-        long_prefill_token_threshold=long_prefill_token_threshold,
-        disable_chunked_mm_input=disable_chunked_mm_input,
-        enable_chunked_prefill=True,
-        async_scheduling=async_scheduling,
-    )
-    model_config = ModelConfig(
-        model=model,
-        task="auto",
-        tokenizer=model,
-        tokenizer_mode="auto",
-        trust_remote_code=True,
-        dtype="float16",
-        seed=42,
-        skip_tokenizer_init=skip_tokenizer_init,
-    )
-    # Cache config, optionally force APC
-    kwargs_cache = ({} if enable_prefix_caching is None else {
-        'enable_prefix_caching': enable_prefix_caching
-    })
-    cache_config = CacheConfig(
-        block_size=block_size,
-        gpu_memory_utilization=0.9,
-        swap_space=0,
-        cache_dtype="auto",
-        **kwargs_cache,
-    )
-    kv_transfer_config = KVTransferConfig(
-        kv_connector="SharedStorageConnector",
-        kv_role="kv_both",
-        kv_connector_extra_config={"shared_storage_path": "local_storage"},
-    ) if use_kv_connector else None
-
-    speculative_config: Optional[SpeculativeConfig] = None
-    if num_speculative_tokens is not None:
-        speculative_config = SpeculativeConfig(
-            model="ngram", num_speculative_tokens=num_speculative_tokens)
-
-    vllm_config = VllmConfig(
-        scheduler_config=scheduler_config,
-        model_config=model_config,
-        cache_config=cache_config,
-        kv_transfer_config=kv_transfer_config,
-        speculative_config=speculative_config,
-    )
-    kv_cache_config = KVCacheConfig(
-        num_blocks=num_blocks,  # A large number of blocks to hold all requests
-        kv_cache_tensors=[],
-        kv_cache_groups=[
-            KVCacheGroupSpec(['layer'],
-                             FullAttentionSpec(block_size, 1, 1, torch.float32,
-                                               False))
-        ],
-    )
-    cache_config.num_gpu_blocks = num_blocks
-    scheduler_cls = AsyncScheduler if async_scheduling else Scheduler
-    return scheduler_cls(
-        vllm_config=vllm_config,
-        kv_cache_config=kv_cache_config,
-        log_stats=True,
-        structured_output_manager=StructuredOutputManager(vllm_config),
-    )
-
-
-def create_requests(
-    num_requests: int,
-    num_tokens: int = 10,
-    mm_positions: Optional[list[PlaceholderRange]] = None,
-    max_tokens: int = 16,
-    stop_token_ids: Optional[list[int]] = None,
-    prompt_logprobs: Optional[int] = None,
-    same_prompt: bool = False,
-) -> list[Request]:
-    sampling_params = SamplingParams(ignore_eos=False,
-                                     max_tokens=max_tokens,
-                                     stop_token_ids=stop_token_ids,
-                                     prompt_logprobs=prompt_logprobs)
-    requests = []
-    for i in range(num_requests):
-        if mm_positions is not None:
-            mm_position = mm_positions[i]
-            mm_inputs = [MultiModalKwargs({})] * len(mm_position)
-        else:
-            mm_position = None
-            mm_inputs = None
-        prompt_token_ids = ([0] * num_tokens if same_prompt else [i] *
-                            num_tokens)
-        request = Request(
-            request_id=f"{i}",
-            prompt_token_ids=prompt_token_ids,
-            sampling_params=sampling_params,
-            pooling_params=None,
-            multi_modal_inputs=mm_inputs,
-            multi_modal_placeholders=mm_position,
-            multi_modal_hashes=None,
-            eos_token_id=EOS_TOKEN_ID,
-        )
-        requests.append(request)
-    return requests

tests/v1/e2e/test_spec_decode.py

@@ -6,10 +6,8 @@ import random
 from typing import Any
 
 import pytest
-import torch
 
 from vllm import LLM, SamplingParams
-from vllm.distributed import cleanup_dist_env_and_memory
 
 
 @pytest.fixture
@@ -55,6 +53,14 @@ def model_name():
     return "meta-llama/Llama-3.1-8B-Instruct"
 
 
+def eagle_model_name():
+    return "yuhuili/EAGLE-LLaMA3.1-Instruct-8B"
+
+
+def eagle3_model_name():
+    return "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"
+
+
 def test_ngram_correctness(
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
@@ -71,8 +77,6 @@ def test_ngram_correctness(
         ref_llm = LLM(model=model_name, max_model_len=1024)
         ref_outputs = ref_llm.chat(test_prompts, sampling_config)
         del ref_llm
-        torch.cuda.empty_cache()
-        cleanup_dist_env_and_memory()
 
         spec_llm = LLM(
             model=model_name,
@@ -99,50 +103,34 @@ def test_ngram_correctness(
         # Upon failure, inspect the outputs to check for inaccuracy.
         assert matches > int(0.7 * len(ref_outputs))
         del spec_llm
-        torch.cuda.empty_cache()
-        cleanup_dist_env_and_memory()
 
 
-@pytest.mark.parametrize("model_setup", [
-    ("eagle", "meta-llama/Llama-3.1-8B-Instruct",
-     "yuhuili/EAGLE-LLaMA3.1-Instruct-8B", 1),
-    ("eagle3", "meta-llama/Llama-3.1-8B-Instruct",
-     "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", 1),
-    pytest.param(
-        ("eagle", "meta-llama/Llama-4-Scout-17B-16E-Instruct",
-         "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct", 4),
-        marks=pytest.mark.skip(reason="Skipping due to CI OOM issues")),
-],
-                         ids=["llama3_eagle", "llama3_eagle3", "llama4_eagle"])
+@pytest.mark.parametrize("use_eagle3", [False, True], ids=["eagle", "eagle3"])
 def test_eagle_correctness(
     monkeypatch: pytest.MonkeyPatch,
     test_prompts: list[list[dict[str, Any]]],
     sampling_config: SamplingParams,
-    model_setup: tuple[str, str, str, int],
+    model_name: str,
+    use_eagle3: bool,
 ):
     '''
     Compare the outputs of a original LLM and a speculative LLM
     should be the same when using eagle speculative decoding.
-    model_setup: (method, model_name, eagle_model_name, tp_size)
     '''
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
-        method, model_name, spec_model_name, tp_size = model_setup
 
-        ref_llm = LLM(model=model_name,
-                      max_model_len=2048,
-                      tensor_parallel_size=tp_size)
+        ref_llm = LLM(model=model_name, max_model_len=2048)
         ref_outputs = ref_llm.chat(test_prompts, sampling_config)
         del ref_llm
-        torch.cuda.empty_cache()
-        cleanup_dist_env_and_memory()
 
+        spec_model_name = eagle3_model_name(
+        ) if use_eagle3 else eagle_model_name()
         spec_llm = LLM(
             model=model_name,
             trust_remote_code=True,
-            tensor_parallel_size=tp_size,
             speculative_config={
-                "method": method,
+                "method": "eagle3" if use_eagle3 else "eagle",
                 "model": spec_model_name,
                 "num_speculative_tokens": 3,
                 "max_model_len": 2048,
@@ -164,5 +152,3 @@ def test_eagle_correctness(
         # Upon failure, inspect the outputs to check for inaccuracy.
         assert matches > int(0.66 * len(ref_outputs))
         del spec_llm
-        torch.cuda.empty_cache()
-        cleanup_dist_env_and_memory()

tests/v1/entrypoints/openai/responses/test_image.py

@@ -1,166 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import json
-
-import openai
-import pytest
-import pytest_asyncio
-
-from tests.utils import RemoteOpenAIServer
-from vllm.multimodal.utils import encode_image_base64, fetch_image
-
-# Use a small vision model for testing
-MODEL_NAME = "Qwen/Qwen2.5-VL-3B-Instruct"
-MAXIMUM_IMAGES = 2
-# Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
-TEST_IMAGE_URLS = [
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
-    "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
-    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
-    "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
-]
-
-
-@pytest.fixture(scope="module")
-def default_image_server_args():
-    return [
-        "--enforce-eager",
-        "--max-model-len",
-        "6000",
-        "--max-num-seqs",
-        "128",
-        "--limit-mm-per-prompt",
-        json.dumps({"image": MAXIMUM_IMAGES}),
-    ]
-
-
-@pytest.fixture(scope="module")
-def image_server(default_image_server_args):
-    with RemoteOpenAIServer(MODEL_NAME,
-                            default_image_server_args) as remote_server:
-        yield remote_server
-
-
-@pytest_asyncio.fixture
-async def client(image_server):
-    async with image_server.get_async_client() as async_client:
-        yield async_client
-
-
-@pytest.fixture(scope="session")
-def base64_encoded_image() -> dict[str, str]:
-    return {
-        image_url: encode_image_base64(fetch_image(image_url))
-        for image_url in TEST_IMAGE_URLS
-    }
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
-async def test_single_chat_session_image(client: openai.AsyncOpenAI,
-                                         model_name: str, image_url: str):
-    content_text = "What's in this image?"
-    messages = [{
-        "role":
-        "user",
-        "content": [
-            {
-                "type": "input_image",
-                "image_url": image_url,
-                "detail": "auto",
-            },
-            {
-                "type": "input_text",
-                "text": content_text
-            },
-        ],
-    }]
-
-    # test image url
-    response = await client.responses.create(
-        model=model_name,
-        input=messages,
-    )
-    assert len(response.output_text) > 0
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
-async def test_single_chat_session_image_base64encoded(
-    client: openai.AsyncOpenAI,
-    model_name: str,
-    image_url: str,
-    base64_encoded_image: dict[str, str],
-):
-    content_text = "What's in this image?"
-    messages = [{
-        "role":
-        "user",
-        "content": [
-            {
-                "type": "input_image",
-                "image_url":
-                f"data:image/jpeg;base64,{base64_encoded_image[image_url]}",
-                "detail": "auto",
-            },
-            {
-                "type": "input_text",
-                "text": content_text
-            },
-        ],
-    }]
-    # test image base64
-    response = await client.responses.create(
-        model=model_name,
-        input=messages,
-    )
-    assert len(response.output_text) > 0
-
-
-@pytest.mark.asyncio
-@pytest.mark.parametrize("model_name", [MODEL_NAME])
-@pytest.mark.parametrize(
-    "image_urls",
-    [TEST_IMAGE_URLS[:i] for i in range(2, len(TEST_IMAGE_URLS))])
-async def test_multi_image_input(client: openai.AsyncOpenAI, model_name: str,
-                                 image_urls: list[str]):
-    messages = [{
-        "role":
-        "user",
-        "content": [
-            *({
-                "type": "input_image",
-                "image_url": image_url,
-                "detail": "auto",
-            } for image_url in image_urls),
-            {
-                "type": "input_text",
-                "text": "What's in this image?"
-            },
-        ],
-    }]
-
-    if len(image_urls) > MAXIMUM_IMAGES:
-        with pytest.raises(openai.BadRequestError):  # test multi-image input
-            await client.responses.create(
-                model=model_name,
-                input=messages,
-            )
-        # the server should still work afterwards
-        response = await client.responses.create(
-            model=model_name,
-            input=[{
-                "role": "user",
-                "content": "What's the weather like in Paris today?",
-            }],
-        )
-        assert len(response.output_text) > 0
-    else:
-        response = await client.responses.create(
-            model=model_name,
-            input=messages,
-        )
-        assert len(response.output_text) > 0

tests/v1/test_external_lb_dp.py

@@ -17,7 +17,7 @@ MODEL_NAME = "ibm-research/PowerMoE-3b"
 
 # Number of data parallel ranks for external LB testing
 DP_SIZE = int(os.getenv("DP_SIZE", "2"))
-# Default tensor parallel size to use
+# Default tensor parallell size to use
 TP_SIZE = int(os.getenv("TP_SIZE", "1"))
 
 

tests/v1/test_utils.py

@@ -3,7 +3,7 @@
 
 import torch
 
-from vllm.v1.worker.utils import bind_kv_cache
+from vllm.v1.utils import bind_kv_cache
 
 
 def test_bind_kv_cache():