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

perf: disable replay buffer VRAM auto-sizing — cap at 500K entries

Browse Source 
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) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-19 09:20:15 +02:00
parent 68ef8eea3e
commit 6dd1aeca7e
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
config/gpu/h100.toml
View File

@@ -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