xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-19 02:37:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

fixes

Browse Source 
Signed-off-by: Tyler Michael Smith <tysmith@redhat.com>
This commit is contained in:
Tyler Michael Smith 2025-06-20 14:40:21 +00:00
parent d5f206767c
commit 26fd8ca33c
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py
View File

@ -105,7 +105,7 @@ def silu_mul_fp8_quant_deep_gemm(
y: torch.Tensor, # (E, T, 2*H) float32 y: torch.Tensor, # (E, T, 2*H) float32
tokens_per_expert: torch.Tensor, # (E,) number of valid tokens per expert tokens_per_expert: torch.Tensor, # (E,) number of valid tokens per expert
group_size: int = 128, group_size: int = 128,
eps: float = 1e-6, eps: float = 1e-10,
): ):
"""Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales
@ -152,7 +152,7 @@ def silu_mul_fp8_quant_deep_gemm(
f_info = torch.finfo(fp8_dtype) f_info = torch.finfo(fp8_dtype)
fp8_max = f_info.max fp8_max = f_info.max
fp8_min = -f_info.max fp8_min = f_info.min
_silu_mul_fp8_quant_deep_gemm[grid]( _silu_mul_fp8_quant_deep_gemm[grid](
y, y,