youngkingdom/vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[V1] Remove V0 code paths for Hybrid models (#25400)

Browse Source 
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2025-09-23 17:26:13 +02:00
committed by GitHub
parent 2c58742dff
commit a903669e10
31 changed files with 352 additions and 2296 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
tests/models/language/generation/test_hybrid.py
View File

@ -20,7 +20,9 @@ pytestmark = pytest.mark.hybrid_model
SSM_MODELS = [
"state-spaces/mamba-130m-hf",
"tiiuae/falcon-mamba-tiny-dev",
"yujiepan/mamba2-codestral-v0.1-tiny-random",
# mamba2-codestral in transformers is broken pending:
# https://github.com/huggingface/transformers/pull/40861
#"yujiepan/mamba2-codestral-v0.1-tiny-random",
]
HYBRID_MODELS = [
@ -31,18 +33,7 @@ HYBRID_MODELS = [
"ibm-granite/granite-4.0-tiny-preview",
"tiiuae/Falcon-H1-0.5B-Base",
"LiquidAI/LFM2-1.2B",
]
V1_SUPPORTED_MODELS = [
"state-spaces/mamba-130m-hf",
"ai21labs/Jamba-tiny-dev",
"pfnet/plamo-2-1b",
"yujiepan/mamba2-codestral-v0.1-tiny-random",
"Zyphra/Zamba2-1.2B-instruct",
"hmellor/tiny-random-BambaForCausalLM",
"ibm-granite/granite-4.0-tiny-preview",
"tiiuae/Falcon-H1-0.5B-Base",
"LiquidAI/LFM2-1.2B",
"tiny-random/qwen3-next-moe",
]
FULL_CUDA_GRAPH_MODELS = [
@ -51,10 +42,6 @@ FULL_CUDA_GRAPH_MODELS = [
"Zyphra/Zamba2-1.2B-instruct",
]
V0_UNSUPPORTED_MODELS = [
"LiquidAI/LFM2-1.2B",
]
FP32_STATE_MODELS = [
"state-spaces/mamba-130m-hf",
"Zyphra/Zamba2-1.2B-instruct",
@ -88,20 +75,16 @@ def test_models(
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)
if model in V1_SUPPORTED_MODELS:
with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
else:
vllm_v1_outputs = None
with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
if model in V1_SUPPORTED_MODELS:
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_v1_outputs,
name_0="hf",
name_1="vllm-v1",
)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", [SSM_MODELS[0], HYBRID_MODELS[0]])
@ -299,14 +282,14 @@ def test_full_cuda_graph(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_v1_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm-v1",
name_1="vllm",
)
@ -340,12 +323,12 @@ def test_fp32_cache_state(
with vllm_runner(model,
max_num_seqs=MAX_NUM_SEQS,
**{cache_dtype_param: "float32"}) as vllm_model:
vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_v1_outputs,
outputs_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm-v1",
name_1="vllm",
)

13
tests/models/registry.py
View File

@ -312,14 +312,12 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
"PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
"Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
"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",
trust_remote_code=True),
max_transformers_version="4.55.4",
transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers", # noqa: E501
trust_remote_code=True),
"QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
max_transformers_version="4.53",
transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers", # noqa: E501
@ -330,7 +328,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
"Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
"Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
"Qwen3NextForCausalLM": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
min_transformers_version="4.56.2"),
extras={"tiny-random": "tiny-random/qwen3-next-moe"}, # noqa: E501
min_transformers_version="4.56.3"),
"RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
"SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
trust_remote_code=True,
@ -644,7 +643,7 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
trust_remote_code=True,
speculative_model="XiaomiMiMo/MiMo-7B-RL"),
"Qwen3NextMTP": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
min_transformers_version="4.56.2"),
min_transformers_version="4.56.3"),
}
_TRANSFORMERS_BACKEND_MODELS = {