gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: CUDA context leak (4 cubin loads → 1) + PPO test assertion

Browse Source 
Fixed CUDA module leak: experience collector loaded EXPERIENCE_KERNELS_CUBIN
4 separate times to extract different kernel functions. Now loads ONCE
and extracts all 4 functions from the single module.

Fixed PPO test: total_experiences assertion updated (PPO doesn't have
counterfactual doubling, unlike DQN).

Test results:
- ml: 893 passed, 2 failed (database only), 17 ignored
- ml-dqn: 359 passed, 0 failed
- Smoke tests: 9/11 passed (2 fail: 50-epoch gradient collapse on
  smoketest tiny network with full generalization — config issue,
  not code bug. 3-epoch smoke test passes with Sharpe 9.34)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-03-30 13:14:28 +02:00
parent e9062edecf
commit 79a9948dcf
2 changed files with 14 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
View File

@@ -974,21 +974,14 @@ impl GpuExperienceCollector {
let exp_h_b2_f32 = stream.alloc_zeros::<f32>(alloc_episodes * network_dims.3)
.map_err(|e| MLError::ModelError(format!("alloc exp_h_b2_f32: {e}")))?;
// GPU-native domain randomization kernels
let domain_rand_starts_kernel = {
let context = stream.context();
let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
.map_err(|e| MLError::ModelError(format!("dr cubin: {e}")))?;
module.load_function("domain_rand_episode_starts")
.map_err(|e| MLError::ModelError(format!("domain_rand_starts load: {e}")))?
};
let domain_rand_params_kernel = {
let context = stream.context();
let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
.map_err(|e| MLError::ModelError(format!("dr cubin2: {e}")))?;
module.load_function("domain_rand_sim_params")
.map_err(|e| MLError::ModelError(format!("domain_rand_params load: {e}")))?
};
// GPU-native domain randomization + saboteur kernels (ONE cubin load)
let exp_module_extra = stream.context()
.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
.map_err(|e| MLError::ModelError(format!("exp cubin extra: {e}")))?;
let domain_rand_starts_kernel = exp_module_extra.load_function("domain_rand_episode_starts")
.map_err(|e| MLError::ModelError(format!("domain_rand_starts load: {e}")))?;
let domain_rand_params_kernel = exp_module_extra.load_function("domain_rand_sim_params")
.map_err(|e| MLError::ModelError(format!("domain_rand_params load: {e}")))?;
let domain_rand_params_buf = stream.alloc_zeros::<f32>(alloc_episodes * 5)
.map_err(|e| MLError::ModelError(format!("alloc dr_params: {e}")))?;
@@ -1004,21 +997,11 @@ impl GpuExperienceCollector {
.map_err(|e| MLError::ModelError(format!("alloc saboteur_best: {e}")))?;
let saboteur_best_return_buf = stream.alloc_zeros::<f32>(1)
.map_err(|e| MLError::ModelError(format!("alloc saboteur_best_return: {e}")))?;
// Load saboteur kernels from experience cubin
let saboteur_generate_kernel = {
let context = stream.context();
let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
.map_err(|e| MLError::ModelError(format!("saboteur cubin: {e}")))?;
module.load_function("saboteur_generate_params")
.map_err(|e| MLError::ModelError(format!("saboteur_generate load: {e}")))?
};
let saboteur_select_kernel = {
let context = stream.context();
let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
.map_err(|e| MLError::ModelError(format!("saboteur cubin2: {e}")))?;
module.load_function("saboteur_select_best")
.map_err(|e| MLError::ModelError(format!("saboteur_select load: {e}")))?
};
// Saboteur kernels from same module (no extra cubin load)
let saboteur_generate_kernel = exp_module_extra.load_function("saboteur_generate_params")
.map_err(|e| MLError::ModelError(format!("saboteur_generate load: {e}")))?;
let saboteur_select_kernel = exp_module_extra.load_function("saboteur_select_best")
.map_err(|e| MLError::ModelError(format!("saboteur_select load: {e}")))?;
// #31 Load bottleneck tanh+concat kernel from utility cubin (if active)
let bn_tanh_concat_fn = if bn_dim_from_params > 0 {

2
crates/ml/src/cuda_pipeline/mod.rs
View File

@@ -1107,7 +1107,7 @@ mod tests {
let cfg = PpoCollectorConfig::default();
assert_eq!(cfg.n_episodes, 128);
assert_eq!(cfg.timesteps_per_episode, 500);
assert_eq!(cfg.total_experiences(), 128_000); // 2x counterfactual
assert_eq!(cfg.total_experiences(), 64_000); // PPO: no counterfactual doubling
assert!((cfg.gamma - 0.99).abs() < f32::EPSILON);
assert!((cfg.gae_lambda - 0.95).abs() < f32::EPSILON);
}