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

fix: L1 output TMA smem_inner_dim was block_n/4, should be block_n/2

Browse Source 
Packed E2M1 output has 2 elements per byte, so block_n elements = block_n/2 bytes.
block_n/4 was under-sizing the TMA SMEM row by 2x → OOB write → LAUNCH_FAILED.
This commit is contained in:
biondizzle
2026-05-12 14:58:11 +00:00
parent d8ae7a3225
commit c71fb97687
2 changed files with 99 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/4 bytes, no swizzle (v1)
// L1 output: packed E2M1, K-dim = intermediate_hidden/2, inner = block_n/2 bytes (packed), 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 / 4, config.store_block_m,
config.block_n / 2, config.store_block_m,
static_cast<int>(l2_acts.stride(-2)),
0, 0, // no swizzle
false, // allow_tf32