diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index 795e691af..ec01617cc 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -3090,14 +3090,6 @@ pub struct GpuDqnTrainer {
     denoise_bias_grad_p1_kernel: CudaFunction,
     /// Kernel handle: denoise_bias_grad_p2.
     denoise_bias_grad_p2_kernel: CudaFunction,
-    /// Dedicated CUfunction handles for OFI bias-grad reduction.
-    /// Loaded from a SEPARATE CUmodule from the denoise variant — each
-    /// child graph (post_aux, adam_grad) must hold its own CUfunction
-    /// reference, otherwise the driver's per-node handle validation on
-    /// Ada/Hopper trips CUDA_ERROR_ILLEGAL_ADDRESS at kernel launch
-    /// (same pattern as graph_safe_copy_kernel below).
-    ofi_embed_bias_grad_p1_kernel: CudaFunction,
-    ofi_embed_bias_grad_p2_kernel: CudaFunction,
 
     // ── xLSTM temporal context (mLSTM cell) ────────────────────────────
     /// [528] 6 weight matrices × [11, 8]: W_k, W_v, W_q, W_i, W_f, W_o
@@ -5188,7 +5180,7 @@ impl GpuDqnTrainer {
         let grad_b_ptr = self.ofi_embed_grad_b.raw_ptr();
         let dim_i32 = OFI_EMBED_DIM as i32;
         unsafe {
-            self.stream.launch_builder(&self.ofi_embed_bias_grad_p1_kernel)
+            self.stream.launch_builder(&self.denoise_bias_grad_p1_kernel)
                 .arg(&d_combined_ptr)
                 .arg(&partials_ptr)
                 .arg(&dim_i32)
@@ -5201,7 +5193,7 @@ impl GpuDqnTrainer {
                 .map_err(|e| MLError::ModelError(format!("ofi_embed_bias_grad_p1: {e}")))?;
         }
         unsafe {
-            self.stream.launch_builder(&self.ofi_embed_bias_grad_p2_kernel)
+            self.stream.launch_builder(&self.denoise_bias_grad_p2_kernel)
                 .arg(&partials_ptr)
                 .arg(&grad_b_ptr)
                 .arg(&dim_i32)
@@ -11105,16 +11097,6 @@ impl GpuDqnTrainer {
             .map_err(|e| MLError::ModelError(format!("denoise_bias_grad_p1 load: {e}")))?;
         let denoise_bias_grad_p2_kernel = cpbi_module_denoise_cublas.load_function("denoise_bias_grad_p2")
             .map_err(|e| MLError::ModelError(format!("denoise_bias_grad_p2 load: {e}")))?;
-        // Dedicated CUmodule for OFI bias-grad reduction (Ada/Hopper child-graph
-        // CUfunction isolation — same pattern as graph_safe_copy below).
-        let ofi_embed_bias_grad_module = stream.context().load_cubin(EXPECTED_Q_CUBIN.to_vec())
-            .map_err(|e| MLError::ModelError(format!("ofi_embed_bias_grad module load: {e}")))?;
-        let ofi_embed_bias_grad_p1_kernel = ofi_embed_bias_grad_module
-            .load_function("denoise_bias_grad_p1")
-            .map_err(|e| MLError::ModelError(format!("ofi_embed_bias_grad_p1 load: {e}")))?;
-        let ofi_embed_bias_grad_p2_kernel = ofi_embed_bias_grad_module
-            .load_function("denoise_bias_grad_p2")
-            .map_err(|e| MLError::ModelError(format!("ofi_embed_bias_grad_p2 load: {e}")))?;
 
         // cuBLAS GEMM descriptors for denoise MLP backward
         let denoise_lt_handle = shared_cublas.lt_handle.0;
@@ -12055,8 +12037,6 @@ impl GpuDqnTrainer {
             denoise_skip_conn_kernel,
             denoise_bias_grad_p1_kernel,
             denoise_bias_grad_p2_kernel,
-            ofi_embed_bias_grad_p1_kernel,
-            ofi_embed_bias_grad_p2_kernel,
             qlstm_weights,
             qlstm_c,
             qlstm_n,
diff --git a/docs/dqn-wire-up-audit.md b/docs/dqn-wire-up-audit.md
index d1dd14685..e04d13b73 100644
--- a/docs/dqn-wire-up-audit.md
+++ b/docs/dqn-wire-up-audit.md
@@ -858,15 +858,3 @@ architecture pretended to do. Workspace clean: 0 errors across all crates, tests
 Smoke validates no regression: 3/3 folds passed (405s), gate utilization
 preserved (util=0.119,0.119,0.169,...), HEALTH_DIAG aux_moe line emits,
 all fold checkpoints written.
-
-OFI bias-grad CUfunction isolation (2026-04-27): fixed
-CUDA_ERROR_ILLEGAL_ADDRESS crash in launch_ofi_embed_backward on L40S
-(workflow train-multi-seed-tdjtr, fold 0). Root cause: shared
-CUfunction handles for denoise_bias_grad_p1/p2 across two captured
-child graphs (post_aux, adam_grad) trip Ada/Hopper driver per-node
-handle validation when the partials buffers land far apart in VRAM.
-Pre-existing latent bug, surfaced by MoE allocator-state shift.
-Fix: dedicated CUmodule for OFI bias-grad kernels, mirroring the
-graph_safe_copy_kernel precedent (line ~11523). Kernel bytecode
-identical; constructor adds ~2ms; zero runtime cost. Smoke validates
-no regression at B=64.