gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

diag: disable multi-stream branch dispatch in f32 forward to test H100 cublasLtMatmul

Browse Source 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-10 13:29:04 +02:00
parent 06f9819f05
commit 45ff940e4b
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
crates/ml/src/cuda_pipeline/batched_forward.rs
View File

@@ -501,9 +501,9 @@ impl CublasForward {
let distinct_branches = h_b0_ptr != h_b1_ptr && h_b1_ptr != h_b2_ptr && h_b0_ptr != h_b2_ptr
&& h_b0_ptr != h_b3_ptr && h_b1_ptr != h_b3_ptr && h_b2_ptr != h_b3_ptr;
if distinct_branches {
// Multi-stream branch dispatch (f32 path, experience collection).
// cublasLtMatmul takes stream per-call — no cublasSetStream needed.
if false && distinct_branches {
// Multi-stream branch dispatch DISABLED for H100 cublasLtMatmul debugging.
// cublasLtMatmul returns NOT_SUPPORTED on branch streams on H100.
self.trunk_done_event.record(stream)
.map_err(|e| MLError::ModelError(format!("f32 trunk event record: {e}")))?;