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

revert: block_n/4 was correct (SwiGLU halving × FP4 packing)

Browse Source
This commit is contained in:
biondizzle
2026-05-12 15:04:23 +00:00
parent c71fb97687
commit 94b30dc2bc
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
csrc/jit_kernels/impls/sm100_fp8_nvfp4_mega_moe.hpp
View File

@@ -157,10 +157,10 @@ static void sm100_fp8_nvfp4_mega_moe(
intermediate_hidden * 2, hidden,
config.block_n, kGranK,
num_experts_per_rank, 0);
// L1 output: packed E2M1, K-dim = intermediate_hidden/2, inner = block_n/2 bytes (packed), no swizzle (v1)
// L1 output: packed E2M1, K-dim = intermediate_hidden/2, inner = block_n/4 bytes (SwiGLU halving × FP4 packing), no swizzle (v1)
const auto tensor_map_l1_output = make_tma_2d_desc(l2_acts,
intermediate_hidden / 2, config.num_max_pool_tokens,
config.block_n / 2, config.store_block_m,
config.block_n / 4, config.store_block_m,
static_cast<int>(l2_acts.stride(-2)),
0, 0, // no swizzle
false, // allow_tf32