6dd1aeca7e
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>
15 lines
433 B
TOML
15 lines
433 B
TOML
# H100 PCIe/SXM (80GB VRAM) -- full production
|
||
[training]
|
||
batch_size = 8192
|
||
num_atoms = 52
|
||
buffer_size = 500000
|
||
hidden_dim_base = 256
|
||
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
|
||
gpu_n_episodes = 4096 # 132 SMs × ~31 eps/SM — full H100 utilization
|
||
|
||
[cuda]
|
||
cuda_stack_bytes = 65536 # 64KB
|