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

diag: force a_format/b_format=5 (MXF8F6F4Format::E2M1), re-enable MMA, dump k=0+k=1

Browse Source
This commit is contained in:
biondizzle
2026-05-12 19:06:28 +00:00
parent 3b8aa5fd4d
commit c1cbe488f3
2 changed files with 31 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
csrc/jit_kernels/heuristics/mega_moe.hpp
View File

@@ -228,7 +228,17 @@ static MegaMoEConfig get_mega_moe_config(
num_dispatch_threads, num_non_epilogue_threads, num_epilogue_threads
};
// Print configs for the first time
// Always print block_m for NVFP4 debugging
{
const auto key = fmt::format(
"MegaMoEConfig(num_ranks={}, num_experts={}, hidden={}, intermediate_hidden={}, num_max_tokens_per_rank={}, num_tokens={}, num_topk={})",
num_ranks, num_experts, hidden, intermediate_hidden, num_max_tokens_per_rank, num_tokens, num_topk);
static std::unordered_set<std::string> printed;
if (printed.count(key) == 0) {
std::cout << "[DIAG-HOST] " << key << ": " << config << std::endl;
printed.insert(key);
}
}
if (get_env<int>("DG_JIT_DEBUG") or get_env<int>("DG_PRINT_CONFIGS")) {
const auto key = fmt::format(
"MegaMoEConfig(num_ranks={}, num_experts={}, hidden={}, intermediate_hidden={}, num_max_tokens_per_rank={}, num_tokens={}, num_topk={})",