foxhunt

gitadmin/foxhunt

Fork 0

Commit Graph

Author	SHA1	Message	Date
jgrusewski	68804a1a51	fix(cuda): remove --use_fast_math, eliminate cross-stream NaN race, f32 rewards/dones Three root causes of sporadic NaN during training: 1. --use_fast_math (nvcc) breaks IEEE 754 NaN semantics: fmaxf(NaN,x) returns NaN instead of x, isnan()/isinf() compile to false. Replaced with --ftz=true --fmad=true --prec-div=true --prec-sqrt=true across all 4 build.rs (ml, ml-dqn, ml-ppo, ml-core). 2. Cross-stream race: replay buffer wrote batch data on the device's original stream while the trainer read it on a forked stream. Fixed by passing the forked stream to the DQN agent via agent_device, so all GPU components share a single CUDA stream (zero sync overhead). 3. Rewards/dones stored as bf16 in replay buffer caused done=0xFFFF NaN. Converted entire rewards/dones pipeline to f32: experience collector, replay buffer storage, nstep kernel, loss/grad kernels. Also: - Removed fast_isnan/fast_isinf/fast_isfinite wrappers — standard isnan/isinf/isfinite work correctly without --use_fast_math - Updated dqn-smoketest.toml: lr=1e-4, epsilon=1e-8 (f32 Adam values) - Removed debug printfs from gather kernels - Added curiosity_weight to training profile system - Cleaned up smoke_params() inline overrides 11/11 smoke tests pass, 5/5 stress runs of 50-epoch test pass, 359/359 ml-dqn + 895/895 ml unit tests pass. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-03-29 16:18:37 +02:00
jgrusewski	07d0e60fe4	feat(bf16): remove nvrtc from entire workspace + wire ml-core precompiled cubins - Fork cudarc locally (vendor/cudarc): add CudaContext::load_cubin() that calls cuModuleLoadData directly — zero nvrtc dependency - Remove "nvrtc" feature from ml-core, ml-dqn, ml-ppo Cargo.toml - Replace all 89 Ptx::from_binary + load_module calls with load_cubin - ml-core cuda_autograd: wire 9 stub constructors to precompiled cubins (activation, elementwise, linear, loss, reduction, dropout, layer_norm, optimizer) - ml-core build.rs: compile 8 BF16-native CUDA kernels via nvcc - cubin_loader.rs: thin wrapper around CudaContext::load_cubin() - Fix size_of::<f32> in gpu_tensor.rs, stream_ops.rs, layer_norm.rs - Fix test data: Vec<f32> → Vec<half::bf16> for memcpy_htod - Stub ml-ppo/ml-dqn runtime compile_ptx calls (dead code) - backtest_metrics_kernel.cu: full native BF16 rewrite (no float) - backtest_env_kernel.cu: shared memory → __nv_bfloat16 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-03-28 10:11:46 +01:00

Author

SHA1

Message

Date

jgrusewski

68804a1a51

fix(cuda): remove --use_fast_math, eliminate cross-stream NaN race, f32 rewards/dones

Three root causes of sporadic NaN during training:

1. --use_fast_math (nvcc) breaks IEEE 754 NaN semantics: fmaxf(NaN,x)
   returns NaN instead of x, isnan()/isinf() compile to false.
   Replaced with --ftz=true --fmad=true --prec-div=true --prec-sqrt=true
   across all 4 build.rs (ml, ml-dqn, ml-ppo, ml-core).

2. Cross-stream race: replay buffer wrote batch data on the device's
   original stream while the trainer read it on a forked stream.
   Fixed by passing the forked stream to the DQN agent via agent_device,
   so all GPU components share a single CUDA stream (zero sync overhead).

3. Rewards/dones stored as bf16 in replay buffer caused done=0xFFFF NaN.
   Converted entire rewards/dones pipeline to f32: experience collector,
   replay buffer storage, nstep kernel, loss/grad kernels.

Also:
- Removed fast_isnan/fast_isinf/fast_isfinite wrappers — standard
  isnan/isinf/isfinite work correctly without --use_fast_math
- Updated dqn-smoketest.toml: lr=1e-4, epsilon=1e-8 (f32 Adam values)
- Removed debug printfs from gather kernels
- Added curiosity_weight to training profile system
- Cleaned up smoke_params() inline overrides

11/11 smoke tests pass, 5/5 stress runs of 50-epoch test pass,
359/359 ml-dqn + 895/895 ml unit tests pass.

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

2026-03-29 16:18:37 +02:00

jgrusewski

07d0e60fe4

feat(bf16): remove nvrtc from entire workspace + wire ml-core precompiled cubins

- Fork cudarc locally (vendor/cudarc): add CudaContext::load_cubin()
  that calls cuModuleLoadData directly — zero nvrtc dependency
- Remove "nvrtc" feature from ml-core, ml-dqn, ml-ppo Cargo.toml
- Replace all 89 Ptx::from_binary + load_module calls with load_cubin
- ml-core cuda_autograd: wire 9 stub constructors to precompiled cubins
  (activation, elementwise, linear, loss, reduction, dropout, layer_norm, optimizer)
- ml-core build.rs: compile 8 BF16-native CUDA kernels via nvcc
- cubin_loader.rs: thin wrapper around CudaContext::load_cubin()
- Fix size_of::<f32> in gpu_tensor.rs, stream_ops.rs, layer_norm.rs
- Fix test data: Vec<f32> → Vec<half::bf16> for memcpy_htod
- Stub ml-ppo/ml-dqn runtime compile_ptx calls (dead code)
- backtest_metrics_kernel.cu: full native BF16 rewrite (no float)
- backtest_env_kernel.cu: shared memory → __nv_bfloat16

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

2026-03-28 10:11:46 +01:00

2 Commits