b4cf279b73
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) <noreply@anthropic.com>
15 lines
405 B
TOML
15 lines
405 B
TOML
# H100 PCIe/SXM (80GB VRAM) -- full production
|
||
[training]
|
||
batch_size = 16384
|
||
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
|
||
|
||
[experience]
|
||
gpu_timesteps_per_episode = 5000
|
||
gpu_n_episodes = 4096 # 132 SMs × ~31 eps/SM — full H100 utilization
|
||
|
||
[cuda]
|
||
cuda_stack_bytes = 65536 # 64KB
|