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

[bugfix] Fix static asymmetric quantization case (#10334)

Browse Source 
Signed-off-by: Daniël de Kok <me@danieldk.eu>
Signed-off-by: luka <luka@neuralmagic.com>
Co-authored-by: Daniël de Kok <me@danieldk.eu>
This commit is contained in:
Luka Govedič
2024-11-14 20:35:11 -05:00
committed by GitHub
parent 972112d82f
commit bf2ddc6610
5 changed files with 58 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
tests/quantization/test_compressed_tensors.py
View File

@ -8,6 +8,7 @@ import pytest
import torch
from compressed_tensors.quantization import QuantizationType
from tests.models.utils import check_logprobs_close
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
CompressedTensorsLinearMethod, CompressedTensorsW4A16Sparse24,
CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
@ -74,6 +75,35 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner, model_args):
assert output
@pytest.mark.parametrize(
"model_path",
[
"neuralmagic/Llama-3.2-1B-quantized.w8a8"
# TODO static & asymmetric
])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [10])
def test_compressed_tensors_w8a8_logprobs(hf_runner, vllm_runner,
example_prompts, model_path,
max_tokens, num_logprobs):
dtype = "bfloat16"
with hf_runner(model_path, dtype=dtype) as hf_model:
hf_outputs = hf_model.generate_greedy_logprobs_limit(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(model_path, dtype=dtype) as vllm_model:
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_outputs,
name_0="hf",
name_1="vllm",
)
def test_compressed_tensors_no_enforce_eager(vllm_runner):
model_path = "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change"
with vllm_runner(model_path) as llm: