Files
foxhunt/crates
jgrusewski d83c1d4c48 refactor(bf16): Spec C dead code cleanup — delete 911 LOC of F32 paths
batched_forward.rs (-698 lines):
- Delete sgemm_layer, sgemm_layer_raw (F32 cublasSgemm)
- Delete 4 F32 bias launchers (launch_add_bias_relu/_raw, launch_add_bias/_raw)
- Delete forward_online_bf16, forward_target_bf16 (conversion layer paths)
- Delete bf16_weight_ptrs, 6 dead BF16 buffer accessors, raw_u16_ptr
- Delete 15 CudaSlice<u16> internal BF16 mirror buffers from CublasForward
- Delete F32 kernel fields (add_bias_relu_kernel, add_bias_kernel, f32_to_bf16_kernel)
- compile_bias_kernels returns only BF16 kernels now

gpu_dqn_trainer.rs (-199 lines):
- Delete bf16_params_buf, bf16_target_params_buf mirror infrastructure
- Delete launch_segmented_bf16_convert, launch_bf16_convert_online/target
- Delete bf16_goff_byte_offsets, bf16_padded_total, bf16_mirrors_initialized
- Delete bf16_to_f32_kernel accessor, raw_device_ptr_u16 helper
- Remove sync_target_bf16 call from target_ema_update

fused_training.rs (4 size_of::<f32> → size_of::<half::bf16>):
- td_errors DtoD copy, ensemble buffer memset, dueling/branching weight clones

1722/1722 tests pass. Zero dead F32 code remains.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-28 14:14:31 +01:00
..