[Bugfix][Quantization] Register Quark per-block FP8 scales as weight_scale (#54770)

Signed-off-by: jimmy-adams <41593649+jimmy-adams@users.noreply.github.com>
Co-authored-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
This commit is contained in:
jimmy-adams
2026-09-05 10:44:24 +08:00
committed by GitHub
co-authored by fxmarty-amd Shanshan Shen
parent 6cbb3c154e
commit 385ba6b5b0
7 changed files with 111 additions and 25 deletions
+63
View File
@@ -11,6 +11,7 @@ import importlib.metadata
from dataclasses import dataclass
from importlib.util import find_spec
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
import huggingface_hub
import lm_eval
@@ -702,6 +703,68 @@ def enable_pickle(monkeypatch):
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
def test_quark_w8a8_fp8_per_block_registers_weight_scale(monkeypatch):
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
get_fp8_block_weight_scale,
)
monkeypatch.setattr(
"vllm.model_executor.layers.quantization.quark.schemes."
"quark_w8a8_fp8.get_current_vllm_config",
lambda: SimpleNamespace(model_config=SimpleNamespace(dtype=torch.bfloat16)),
)
scheme = QuarkW8A8Fp8PerBlock(kFp8Static128BlockSym, kFp8Dynamic128Sym)
layer = torch.nn.Module()
layer.weight_scale = torch.tensor([2.0])
assert get_fp8_block_weight_scale(layer) is None
layer.scheme = scheme
assert get_fp8_block_weight_scale(layer) is layer.weight_scale
layer.weight_scale_inv = torch.tensor([3.0])
assert get_fp8_block_weight_scale(layer) is layer.weight_scale
layer.scheme = None
assert get_fp8_block_weight_scale(layer) is layer.weight_scale_inv
loaded = torch.nn.Module()
def weight_loader(param, loaded_weight):
return None
dummy_param = torch.nn.Parameter(torch.empty(1), requires_grad=False)
with (
patch(
"vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8."
"validate_fp8_block_shape"
),
patch(
"vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8."
"create_fp8_weight_parameter",
return_value=dummy_param,
),
patch(
"vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8."
"create_fp8_scale_parameter",
return_value=dummy_param,
),
patch(
"vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8."
"init_fp8_linear_kernel",
return_value=MagicMock(),
),
):
scheme.create_weights(
loaded,
output_partition_sizes=[256],
input_size_per_partition=256,
params_dtype=torch.bfloat16,
weight_loader=weight_loader,
input_size=256,
output_size=256,
)
assert hasattr(loaded, "weight_scale")
assert not hasattr(loaded, "weight_scale_inv")
def test_quark_config_has_no_model_specific_fused_mappings():
config = QuarkConfig({})