6e40e86c09
Hyperopt: - num_atoms: 51-201 → 11-51 (GPU profile caps to hardware limit) - hidden_dim_base: 256-1024 → 128-512 (1024+ is wasteful for 2-layer net) - Prevents wildly oversized networks (num_atoms=200 caused 25s/epoch) H100 experience: - gpu_n_episodes: 256 → 2048 (8x larger cuBLAS batch saturates 132 SMs) - gpu_timesteps_per_episode: 500 → 100 (fewer steps, more parallel episodes) - Total experiences: 204K/epoch (was 128K) with better GPU utilization - Expected: experience 357ms → ~100ms (SM utilization 5% → 40%+) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
18 lines
464 B
TOML
18 lines
464 B
TOML
# H100 PCIe/SXM (80GB VRAM) -- full production
|
||
[training]
|
||
batch_size = 1024
|
||
num_atoms = 51
|
||
buffer_size = 500_000
|
||
hidden_dim_base = 256
|
||
replay_buffer_vram_fraction = 0.70
|
||
|
||
[experience]
|
||
# 2048 episodes: cuBLAS SGEMM batch=2048 saturates 132 SMs.
|
||
# At 256 episodes, the tiny 256×48 GEMMs leave >90% of SMs idle.
|
||
# 2048×100 = 204K experiences/epoch (vs 256×500 = 128K).
|
||
gpu_n_episodes = 2048
|
||
gpu_timesteps_per_episode = 100
|
||
|
||
[cuda]
|
||
cuda_stack_bytes = 65536 # 64KB
|