diff --git a/single_shot_inference.py b/single_shot_inference.py
index 369b1d50..96021c8b 100644
--- a/single_shot_inference.py
+++ b/single_shot_inference.py
@@ -1397,21 +1397,15 @@ def main():
     torch.cuda.set_device(0)
     embed_w = all_w.get("model.embed_tokens.weight")
     embed = torch.nn.Embedding.from_pretrained(embed_w.bfloat16().to('cuda:0'))
-    # lm_head: NVFP4 production GEMM
+    # lm_head: BF16 GEMM — checkpoint weight is BF16, no quantization needed.
+    # NVFP4 quantization of lm_head can flatten/reorder vocabulary logits;
+    # BF16 preserves the full logit distribution for accurate token selection.
     lm_w_raw = all_w.get("lm_head.weight", embed_w).bfloat16().to('cuda:0')
-    from dsv4.layers.linear import Nvfp4Linear
-    lm_head_lin = Nvfp4Linear(lm_w_raw.shape[1], lm_w_raw.shape[0], max_num_tokens=8192, device='cuda:0')
-    from dsv4.ops.quantize import quantize_weight_to_nvfp4
-    lm_fp4, lm_sf, lm_gs = quantize_weight_to_nvfp4(lm_w_raw.T.contiguous())
-    lm_head_lin.fp4 = [lm_fp4.permute(1, 0).contiguous()]
-    lm_head_lin.sf = [lm_sf.permute(1, 0).contiguous()]
-    lm_head_lin.gs = [lm_gs]
-    lm_head_lin.ws2 = [None]
-    lm_head_lin._activation_global_scale = 1.0 / (6.0 * 448.0)
-    lm_head_lin._use_runtime_gsa = True
-    lm_head_lin.finalize_weights()
+    # Store as (vocab, hidden) BF16 — direct cuBLAS GEMM
+    lm_head_weight = lm_w_raw.contiguous()  # (129280, 7168) BF16
+    lm_head_lin = None  # flag: use BF16 path
     lm_w = None
-    print("  lm_head: NVFP4 production GEMM")
+    print("  lm_head: BF16 GEMM (checkpoint weight, no quantization)")
     final_norm_w = all_w.get("model.norm.weight")
     if final_norm_w is not None: final_norm_w = final_norm_w.to('cuda:0', torch.float32)
 
@@ -1660,8 +1654,8 @@ def main():
                         gl._activation_global_scale = fixed_gsa
                         gl._use_runtime_gsa = False
                         n_fixed += 1
-            # lm_head
-            if hasattr(lm_head_lin, '_gsa_buf') and hasattr(lm_head_lin, '_use_runtime_gsa') and lm_head_lin._use_runtime_gsa:
+            # lm_head (skip if BF16 path)
+            if lm_head_lin is not None and hasattr(lm_head_lin, '_gsa_buf') and hasattr(lm_head_lin, '_use_runtime_gsa') and lm_head_lin._use_runtime_gsa:
                 fixed_gsa = lm_head_lin._gsa_buf.item()
                 lm_head_lin._activation_global_scale = fixed_gsa
                 lm_head_lin._use_runtime_gsa = False
@@ -1669,7 +1663,12 @@ def main():
             print(f"  Warmup gsa: fixed {n_fixed} projection gsa values from step 0 (MoE/SE keep runtime gsa)", flush=True)
         x_out = hc_head.forward(X) if hc_head is not None else X[:, 0, :]
         if final_norm_w is not None: x_out = rmsnorm(x_out, final_norm_w)
-        logits = lm_head_lin(x_out)
+        # lm_head: BF16 GEMM (cuBLAS) or NVFP4 GEMM
+        if lm_head_lin is not None:
+            logits = lm_head_lin(x_out)
+        else:
+            # BF16 path: direct matmul with (vocab, hidden) weight
+            logits = torch.nn.functional.linear(x_out, lm_head_weight)
         if profile: torch.cuda.synchronize()
         t_lm = time.perf_counter()
         # Check thinking start token logit on first step