feat: DqnGpuData::upload_slices — contiguous array upload from fxcache

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

crates/ml/src/cuda_pipeline/mod.rs

@@ -298,6 +298,73 @@ impl DqnGpuData {
         })
     }
 
+    /// Upload DQN training data to GPU from separate contiguous arrays.
+    ///
+    /// Accepts features, targets, and OFI as separate slices matching the fxcache
+    /// flat binary layout — no tuple allocation or per-bar heap Vec required.
+    /// OFI is uploaded only when at least one value is non-zero.
+    pub fn upload_slices(
+        features: &[[f64; 42]],
+        targets: &[[f64; 4]],
+        ofi: &[[f64; 8]],
+        stream: &Arc<CudaStream>,
+    ) -> Result<Self, MLError> {
+        let num_bars = features.len();
+        if num_bars == 0 {
+            return Err(MLError::ModelError("Empty training data".to_owned()));
+        }
+
+        let feature_dim = 42;
+        let target_dim = 4;
+
+        let estimated_bytes = estimate_vram_bytes(num_bars * (feature_dim + target_dim));
+        if estimated_bytes > MAX_UPLOAD_BYTES {
+            return Err(MLError::ModelError(format!(
+                "Training data too large for GPU upload: {:.1} GB > 2.0 GB limit ({} bars)",
+                estimated_bytes as f64 / 1_073_741_824.0,
+                num_bars,
+            )));
+        }
+
+        let flat_features: Vec<f32> = features
+            .iter()
+            .flat_map(|f| f.iter().map(|&v| v as f32))
+            .collect();
+
+        let flat_targets: Vec<f32> = targets
+            .iter()
+            .flat_map(|t| t.iter().map(|&v| v as f32))
+            .collect();
+
+        let features_gpu = clone_htod_f32_to_bf16(stream, &flat_features)?;
+        let targets_gpu = clone_htod_f32_to_bf16(stream, &flat_targets)?;
+
+        let ofi_features = if !ofi.is_empty() && ofi.iter().any(|row| row.iter().any(|&v| v != 0.0)) {
+            let mut flat_ofi = Vec::with_capacity(num_bars * 8);
+            for i in 0..num_bars {
+                if i < ofi.len() {
+                    for &v in ofi[i].iter() {
+                        flat_ofi.push(v as f32);
+                    }
+                } else {
+                    flat_ofi.extend_from_slice(&[0.0_f32; 8]);
+                }
+            }
+            Some(clone_htod_f32_to_bf16(stream, &flat_ofi)?)
+        } else {
+            None
+        };
+
+        Ok(Self {
+            features: features_gpu,
+            targets: targets_gpu,
+            ofi_features,
+            num_bars,
+            feature_dim,
+            aligned_state_dim: None,
+        })
+    }
+
     /// Upload OFI features (8 dims per bar) from MBP-10 order book data.
     ///
     /// Must be called after `upload()` with a slice matching `num_bars` length.