iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

fix: resolve CUTLASS fmin compatibility for DeepSeek-V4 init (#44236)

Browse Source 
Signed-off-by: Willow Lopez <100782273+Oxygen56@users.noreply.github.com>
This commit is contained in:
Willow Lopez
2026-06-03 01:07:10 -04:00
committed by GitHub
parent 3f0a91bb96
commit 597bc15936
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/models/deepseek_v4/nvidia/ops/sparse_attn_compress_cutedsl.py
+4 -4
View File
@@ -370,11 +370,11 @@ class SparseAttnCompressNormRopeStoreC4Kernel:
inv_scale = _recast_val((Uint32(254) - ue8m0) << Uint32(23), Float32)
for pair in cutlass.range_constexpr(self.elems_per_lane // 2):
elem = const_expr(pair * 2)
y0 = cute.arch.fmin(
y0 = cutlass.min(
cute.arch.fmax(q[elem] * inv_scale, Float32(-self.fp8_max)),
Float32(self.fp8_max),
)
y1 = cute.arch.fmin(
y1 = cutlass.min(
cute.arch.fmax(q[elem + 1] * inv_scale, Float32(-self.fp8_max)),
Float32(self.fp8_max),
)
@@ -1026,11 +1026,11 @@ class SparseAttnNormRopeStoreKernel:
bits = _recast_val(scale_raw, Uint32)
ue8m0 = ((bits + Uint32(0x7FFFFF)) >> Uint32(23)) & Uint32(0xFF)
inv_scale = _recast_val((Uint32(254) - ue8m0) << Uint32(23), Float32)
y0 = cute.arch.fmin(
y0 = cutlass.min(
cute.arch.fmax(q0 * inv_scale, Float32(-self.fp8_max)),
Float32(self.fp8_max),
)
y1 = cute.arch.fmin(
y1 = cutlass.min(
cute.arch.fmax(q1 * inv_scale, Float32(-self.fp8_max)),
Float32(self.fp8_max),
)