From 45ff940e4bfd922d149fba95208e6e5d3f02863a Mon Sep 17 00:00:00 2001 From: jgrusewski Date: Fri, 10 Apr 2026 13:29:04 +0200 Subject: [PATCH] diag: disable multi-stream branch dispatch in f32 forward to test H100 cublasLtMatmul Co-Authored-By: Claude Opus 4.6 (1M context) --- crates/ml/src/cuda_pipeline/batched_forward.rs | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/crates/ml/src/cuda_pipeline/batched_forward.rs b/crates/ml/src/cuda_pipeline/batched_forward.rs index 0dcc79747..4b5f9debc 100644 --- a/crates/ml/src/cuda_pipeline/batched_forward.rs +++ b/crates/ml/src/cuda_pipeline/batched_forward.rs @@ -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}")))?;