biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[W8A8 Block Linear Refactor][2/N] Remove W8A8Fp8BlockLinearOp and adopt Fp8 block linear kernel selections. (#33892)

Browse Source 
Signed-off-by: maral <maralbahari.98@gmail.com>
Signed-off-by: Maral <maralbahari.98@gmail.com>
This commit is contained in:
Maral
2026-04-09 08:50:39 +08:00
committed by GitHub
parent 8332078cfd
commit 2e9034c998
35 changed files with 1710 additions and 904 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/quantization/test_modelopt.py
View File

@@ -12,6 +12,7 @@ import pytest
import torch
from tests.quantization.utils import is_quant_method_supported
from vllm.config.model import ModelConfig
@pytest.fixture(scope="function", autouse=True)
@@ -46,7 +47,7 @@ def _snapshot_download_or_skip(model_id: str) -> str:
not is_quant_method_supported("modelopt"),
reason="ModelOpt FP8 is not supported on this GPU type.",
)
def test_modelopt_fp8_checkpoint_setup(vllm_runner):
def test_modelopt_fp8_checkpoint_setup(default_vllm_config, vllm_runner):
"""Test ModelOpt FP8 checkpoint loading and structure validation."""
# TODO: provide a small publicly available test checkpoint
model_path = (
@@ -61,6 +62,8 @@ def test_modelopt_fp8_checkpoint_setup(vllm_runner):
"This test requires a local ModelOpt FP8 checkpoint."
)
# Set model config as model_config.dtype is required in ModelOptFp8LinearMethod.
default_vllm_config.model_config = ModelConfig()
with vllm_runner(model_path, quantization="modelopt", enforce_eager=True) as llm:
def check_model(model):
@@ -120,11 +123,13 @@ def test_modelopt_fp8_checkpoint_setup(vllm_runner):
not is_quant_method_supported("modelopt"),
reason="ModelOpt FP8 is not supported on this GPU type.",
)
def test_modelopt_fp8_pc_pt_checkpoint_setup(vllm_runner):
def test_modelopt_fp8_pc_pt_checkpoint_setup(default_vllm_config, vllm_runner):
"""Test ModelOpt FP8_PER_CHANNEL_PER_TOKEN checkpoint setup."""
model_id = "CedricHwang/qwen2.5-0.5b-modelopt-fp8-pc-pt"
model_path = _snapshot_download_or_skip(model_id)
# Set model config as model_config.dtype is required in ModelOptFp8LinearMethod.
default_vllm_config.model_config = ModelConfig()
with vllm_runner(model_path, quantization="modelopt", enforce_eager=True) as llm:
def check_model(model):
@@ -181,11 +186,13 @@ def test_modelopt_fp8_pc_pt_checkpoint_setup(vllm_runner):
not is_quant_method_supported("modelopt"),
reason="ModelOpt FP8 is not supported on this GPU type.",
)
def test_modelopt_fp8_pb_wo_checkpoint_setup(vllm_runner):
def test_modelopt_fp8_pb_wo_checkpoint_setup(default_vllm_config, vllm_runner):
"""Test ModelOpt FP8_PB_WO checkpoint setup."""
model_id = "CedricHwang/qwen2.5-0.5b-modelopt-fp8-pb-wo"
model_path = _snapshot_download_or_skip(model_id)
# Set model config as model_config.dtype is required in ModelOptFp8LinearMethod.
default_vllm_config.model_config = ModelConfig()
with vllm_runner(model_path, quantization="modelopt", enforce_eager=True) as llm:
def check_model(model):