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

fix(ml): pad remaining state tensor paths for tensor core alignment

Browse Source 
Additional padding in get_q_values and convert_to_state — these are
currently unused in the training pipeline but would crash if called.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-03-08 03:24:49 +01:00
parent 2a9e185db9
commit c1287efbcf
1 changed files with 21 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
crates/ml/src/trainers/dqn/trainer.rs
View File

@@ -1264,7 +1264,16 @@ impl DQNTrainer {
async fn get_q_values(&self, state: &TradingState) -> Result<Vec<f64>> {
let agent = self.agent.read().await;
let state_vec = state.to_vector();
let state_tensor = Tensor::new(&*state_vec, &self.device)?.unsqueeze(0)?; // Add batch dimension
let raw_dim = state_vec.len();
let aligned = crate::dqn::mixed_precision::align_dim_for_tensor_cores(raw_dim, &self.device);
let padded: Vec<f32> = if aligned > raw_dim {
let mut v = state_vec.to_vec();
v.resize(aligned, 0.0);
v
} else {
state_vec.to_vec()
};
let state_tensor = Tensor::new(&*padded, &self.device)?.unsqueeze(0)?; // Add batch dimension
let q_values_tensor = agent.forward(&state_tensor)?;
let q_values_vec = q_values_tensor.squeeze(0)?.to_vec1::<f32>()?;
@@ -4221,11 +4230,20 @@ impl DQNTrainer {
// Use internal conversion method (returns TradingState)
let trading_state = self.feature_vector_to_state(feature_vec, Some(close))?;
// Convert TradingState to flat vector
// Convert TradingState to flat vector, pad for tensor core alignment
let state_vec = trading_state.to_vector();
let raw_dim = state_vec.len();
let aligned = crate::dqn::mixed_precision::align_dim_for_tensor_cores(raw_dim, &self.device);
let padded: Vec<f32> = if aligned > raw_dim {
let mut v = state_vec.to_vec();
v.resize(aligned, 0.0);
v
} else {
state_vec.to_vec()
};
// Convert to Tensor using trainer's device (GPU or CPU)
Tensor::new(state_vec.as_slice(), &self.device)
Tensor::new(padded.as_slice(), &self.device)
.context("Failed to create state tensor from TradingState")
}