jgrusewski 709fffe923 fix: revert c51_loss_reduce to (1,1,1) — same Hopper graph hang class ...

The warp-parallel (32,1,1) change to c51_loss_reduce caused the same
hang as isv_feature_gate: changing block_dim inside graph_mega alters
CUDA Graph node structure on Hopper's TMA scheduler. Must stay (1,1,1).

Also: batch_size 8192→16384, gpu_n_episodes 1024→4096, num_atoms 51→52
to align h100.toml with dqn-production.toml.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-17 14:46:15 +02:00

..

a100.toml

fix: restore explicit buffer_size in GPU profiles (auto-sizing removed)

2026-04-04 01:24:19 +02:00

default.toml

fix: restore explicit buffer_size in GPU profiles (auto-sizing removed)

2026-04-04 01:24:19 +02:00

h100.toml

fix: revert c51_loss_reduce to (1,1,1) — same Hopper graph hang class

2026-04-17 14:46:15 +02:00

rtx3050.toml

fix: restore explicit buffer_size in GPU profiles (auto-sizing removed)

2026-04-04 01:24:19 +02:00