[Perf][DSv4.1] Restore the fused query RMSNorm + MXFP8 quantization path (#57679)

Signed-off-by: Juntian Liu <Juntianl777@gmail.com>
This commit is contained in:
Juntian Liu
2026-09-25 01:39:03 +00:00
committed by GitHub
parent 3652f35e7c
commit 609731e8f5
@@ -3,7 +3,10 @@
import torch
from vllm.model_executor.layers.fusion.quant_activation import QuantizedActivation
from vllm.model_executor.layers.fusion.quant_activation import (
QuantizedActivation,
get_input_quant_key,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Dynamic
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
@@ -102,6 +105,7 @@ def fused_q_kv_rmsnorm_quant(
)
if tokens:
block = triton.next_power_of_2(max(q_size, kv_size))
launch_pdl = current_platform.is_arch_support_pdl()
_q_kv_norm_quant_kernel[(padded_tokens, 2)](
qr,
kv,
@@ -117,8 +121,9 @@ def fused_q_kv_rmsnorm_quant(
q_size,
kv_size,
block,
current_platform.is_arch_support_pdl(),
launch_pdl,
num_warps=8 if block >= 2048 else 4,
launch_pdl=launch_pdl,
)
return QuantizedActivation(qo, scales, qr.dtype, qr.shape, kMxfp8Dynamic), kvo
@@ -135,7 +140,7 @@ def can_fuse_query_quant(linears: list[torch.nn.Module]) -> bool:
# QuantKey does not encode scale layout; restrict this producer to the
# consumers that accept F8_128x4 swizzled scales.
return all(
getattr(linear, "input_quant_key", None) == kMxfp8Dynamic
get_input_quant_key(linear) == kMxfp8Dynamic
and type(getattr(getattr(linear, "quant_method", None), "kernel", None))
in (FlashInferCutedslMxfp8LinearKernel, FlashInferCutlassMxfp8LinearKernel)
for linear in linears