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

10
tests/quantization/test_compressed_tensors.py
View File

@@ -16,6 +16,9 @@ from compressed_tensors.quantization import (
)
from tests.models.utils import check_logprobs_close
from vllm.model_executor.kernels.linear import (
Fp8BlockScaledMMLinearKernel,
)
from vllm.model_executor.layers.fused_moe import UnquantizedFusedMoEMethod
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
CompressedTensorsConfig,
@@ -29,7 +32,6 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
CompressedTensorsWNA16,
)
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.fp8_utils import W8A8BlockFp8LinearOp
from vllm.model_executor.layers.quantization.utils.nvfp4_utils import (
cutlass_fp4_supported,
)
@@ -473,16 +475,14 @@ def test_compressed_tensors_fp8_block_enabled(vllm_runner):
qkv_proj = layer.self_attn.qkv_proj
assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8)
assert isinstance(
qkv_proj.scheme.w8a8_block_fp8_linear, W8A8BlockFp8LinearOp
)
assert isinstance(qkv_proj.scheme.fp8_linear, Fp8BlockScaledMMLinearKernel)
assert qkv_proj.weight.dtype is fp8_dtype
assert qkv_proj.weight_scale.dtype is torch.float32
assert len(qkv_proj.weight.shape) == 2
assert len(qkv_proj.weight_scale.shape) == 2
input_quant_op = qkv_proj.scheme.w8a8_block_fp8_linear.input_quant_op
input_quant_op = qkv_proj.scheme.fp8_linear.quant_fp8
assert isinstance(input_quant_op, QuantFP8)
assert input_quant_op._forward_method in (
input_quant_op.forward_cuda,