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

dynamic distpatch of fp8 kernels (#14245)

Browse Source 
Signed-off-by: Jeff Daily <jeff.daily@amd.com>
This commit is contained in:
Jeff Daily
2025-03-11 07:54:56 -07:00
committed by GitHub
parent 08a1a1121d
commit a1c8f3796c
25 changed files with 292 additions and 159 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/attention/backends/mla/common.py
View File

@@ -219,7 +219,7 @@ from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
CompressedTensorsW8A8Fp8)
from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
Fp8LinearGenericOp, current_platform_fp8_dtype, is_fp8)
Fp8LinearGenericOp, is_fp8)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
scaled_quantize)
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
@@ -826,7 +826,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
W_Q_UK, W_Q_UK_scales = scaled_quantize(
W_Q_UK,
self.reqaunt_weight_group_shape,
quant_dtype=current_platform_fp8_dtype)
quant_dtype=current_platform.fp8_dtype())
# For FP8 save the transpose so we can use
# `apply_w8a8_block_fp8_linear` directly
self.W_Q_UK = W_Q_UK.T.contiguous()
@@ -843,7 +843,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
W_UV_O, W_UV_O_scales = scaled_quantize(
W_UV_O,
self.reqaunt_weight_group_shape,
quant_dtype=current_platform_fp8_dtype)
quant_dtype=current_platform.fp8_dtype())
# For FP8 save the transpose so we can use
# `apply_w8a8_block_fp8_linear` directly
self.W_UV_O = W_UV_O.T.contiguous()