From b4cf279b739f4bb60e1f7eaaba2faf65fb59e9af Mon Sep 17 00:00:00 2001 From: jgrusewski Date: Sat, 18 Apr 2026 18:42:06 +0200 Subject: [PATCH] =?UTF-8?q?fix:=20reduce=20H100=20replay=20buffer=20VRAM?= =?UTF-8?q?=20fraction=2055%=20=E2=86=92=2045%=20=E2=80=94=20OOM=20from=20?= =?UTF-8?q?IQN=208.6GB=20+=20Phase=202?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit IQN with 64 quantiles × batch=16384 allocates 8.6GB for tiled intermediates. Combined with attention (250MB), Phase 2 aux workspaces (64MB), backward branch scratch (16MB), the 55% replay fraction left only 349MB free — not enough for experience collector's 192MB next_states allocation. 45% frees ~8GB headroom: replay buffer 19.4M → ~15.9M (still 60× batch). 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 5875b9d39..0c63ad7bb 100644 --- a/config/gpu/h100.toml +++ b/config/gpu/h100.toml @@ -4,7 +4,7 @@ batch_size = 16384 num_atoms = 52 buffer_size = 500000 hidden_dim_base = 256 -replay_buffer_vram_fraction = 0.55 # state_dim=96, 86 weight tensors — leaves ~33GB for compute +replay_buffer_vram_fraction = 0.45 # state_dim=96, 86 tensors + IQN 8.6GB + attention 250MB + Phase 2 128MB [experience] gpu_timesteps_per_episode = 5000