biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

D1: Tiny 4-mode sP placeholder for TMEM-P path

Browse Source
This commit is contained in:
biondizzle
2026-05-24 04:03:28 +00:00
parent 8d565efd7a
commit 317d3209ac
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
dsv4/kernels/attention/fmha.py
View File

@@ -154,9 +154,12 @@ class FmhaKernel:
sV = smem.allocate_tensor(element_type=self.q_dtype,layout=v_smem_s.outer,byte_alignment=128,swizzle=v_smem_s.inner)
sC = smem.allocate_tensor(element_type=self.o_dtype,layout=c_smem_s.outer,byte_alignment=128,swizzle=c_smem_s.inner)
# sP only needed for SMEM-P path. Save SMEM by allocating tiny buffer for TMEM-P.
_p_alloc_layout = p_smem_s.outer if self.use_smem_p else cute.make_layout((1,))
_p_alloc_swizzle = p_smem_s.inner if self.use_smem_p else cute.make_layout((1,))
sP = smem.allocate_tensor(element_type=self.q_dtype,layout=_p_alloc_layout,byte_alignment=128,swizzle=_p_alloc_swizzle)
if self.use_smem_p:
sP = smem.allocate_tensor(element_type=self.q_dtype,layout=p_smem_s.outer,byte_alignment=128,swizzle=p_smem_s.inner)
else:
# Tiny placeholder that matches the 4-mode slice pattern
_tiny_p_layout = cute.make_layout(((1,1),1,(1,1),1))
sP = smem.allocate_tensor(element_type=self.q_dtype,layout=_tiny_p_layout,byte_alignment=16)
gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))