904185004c
argo-train.sh now auto-selects cuda-compute-cap based on --gpu-pool: - ci-training-h100* → sm_90 (Hopper) - ci-training-l40s → sm_89 (Ada Lovelace) Added config/gpu/l40s.toml: - batch_size=4096 (between H100's 8192 and A100's 2048) - buffer_size=300K (scaled for 48GB VRAM) - gpu_timesteps_per_episode=2000 (bandwidth-limited) - gpu_n_episodes=2048 (scaled from H100's 4096) GPU profile loader maps "L40S" → "l40s" (was "a100" fallback). Also fixed pre-existing test drift: num_atoms=52 in h100.toml/a100.toml was 51 in test expectations (padding alignment for C51 kernels). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
14 lines
231 B
TOML
14 lines
231 B
TOML
# A100 (40-80GB VRAM)
|
|
[training]
|
|
batch_size = 2048
|
|
num_atoms = 52
|
|
buffer_size = 200000
|
|
hidden_dim_base = 256
|
|
replay_buffer_vram_fraction = 0.70
|
|
|
|
[experience]
|
|
gpu_timesteps_per_episode = 500
|
|
|
|
[cuda]
|
|
cuda_stack_bytes = 65536 # 64KB
|