From 6dd1aeca7eb09cdf06dfe2d3f3cd228dd05b98ac Mon Sep 17 00:00:00 2001 From: jgrusewski Date: Sun, 19 Apr 2026 09:20:15 +0200 Subject: [PATCH] =?UTF-8?q?perf:=20disable=20replay=20buffer=20VRAM=20auto?= =?UTF-8?q?-sizing=20=E2=80=94=20cap=20at=20500K=20entries?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Auto-sizer inflated replay buffer to 15.8M entries on H100 (45% of 80GB VRAM). The PER prefix scan inside the CUDA graph processes ALL 15.8M entries every step, dominating step time. With buffer_size=500K, the scan processes 500K entries instead. Co-Authored-By: Claude Opus 4.6 (1M context) --- config/gpu/h100.toml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/config/gpu/h100.toml b/config/gpu/h100.toml index 9ca38f725..6c63ca852 100644 --- a/config/gpu/h100.toml +++ b/config/gpu/h100.toml @@ -4,7 +4,7 @@ batch_size = 8192 num_atoms = 52 buffer_size = 500000 hidden_dim_base = 256 -replay_buffer_vram_fraction = 0.45 # state_dim=96, 86 tensors + IQN 8.6GB + attention 250MB + Phase 2 128MB +replay_buffer_vram_fraction = 0.0 # disabled — use exact buffer_size=500K (was 0.45 → 15.8M entries = slow PER prefix scan in graph) [experience] gpu_timesteps_per_episode = 5000