diff --git a/single_shot_inference.py b/single_shot_inference.py
index 9e6e2ffd..974ec96e 100644
--- a/single_shot_inference.py
+++ b/single_shot_inference.py
@@ -246,13 +246,59 @@ def forward_layer(x, w, li, cfg, rope_cos, rope_sin):
     attn_out = attn_out.permute(1, 0, 2).reshape(T, n_h * hd)  # (T, n_h*hd)
     
     # ---- Output projection: o_a (BF16 grouped) → o_b (NVFP4) ----
-    oa_w = w[f"{pre}.o_a_proj.weight"]  # (n_h*hd_per_group, o_rank) BF16
+    # o_a_proj: grouped linear — input is (T, n_h*hd) reshaped as (T*o_groups, heads_per_group*hd)
+    # Each group: (heads_per_group * hd) → o_lora_rank
+    # Then concatenate: (T*o_groups * o_rank) → o_b → (T, H)
+    oa_w = w[f"{pre}.o_a_proj.weight"]  # BF16
     ob_w = w[f"{pre}.o_b_proj.weight"]
     ob_s = w[f"{pre}.o_b_proj.weight_scale"]
     ob_s2 = w[f"{pre}.o_b_proj.weight_scale_2"]
     
-    # o_a is BF16 grouped linear — treat as dense for baseline
-    grouped = bf16_linear(attn_out, oa_w.cuda())  # (1, o_groups*o_rank)
+    heads_per_group = n_h // o_groups  # 128/16 = 8
+    group_input_dim = heads_per_group * hd  # 8 * 512 = 4096
+    
+    # Reshape attention output for grouped projection
+    # attn_out: (T, n_h * hd) → (T, o_groups, heads_per_group * hd) → (T*o_groups, group_input_dim)
+    attn_grouped = attn_out.reshape(T, o_groups, group_input_dim).reshape(T * o_groups, group_input_dim)
+    
+    # o_a: (o_rank, group_input_dim) per group → total (o_groups * o_rank, o_groups * group_input_dim)
+    # But in the checkpoint it's stored as (o_groups * o_rank, n_h * hd) or similar
+    # The actual shape tells us: (4096, 16384) BF16
+    # 4096 = o_rank * o_groups? No, o_rank=1024, o_groups=16 → 16384
+    # 16384 = n_h * hd / 4? No.
+    # Let's just check: 4096 output, 16384 input
+    # 16384 = 4 * 4096 = ??? That doesn't match n_h*hd=65536
+    
+    # Actually: o_a_proj weight is (4096, 16384). Since it's BF16, no FP4 packing.
+    # So actual out_features=4096, in_features=16384.
+    # 16384 = ??? Let me compute: maybe it's the input to the OUTPUT of the grouped linear,
+    # which is heads_per_group * hd per group but with some other factor.
+    # Actually 16384 = 32 * 512 = 32 heads * hd? That would be 1/4 of the heads.
+    # Or: 16384 = o_groups * (heads_per_group * hd) / something
+    
+    # The simplest approach: just try both possible reshapes
+    try:
+        # Try 1: treat as standard linear with the full attention output
+        if oa_w.shape[1] * 1 == n_h * hd:  # BF16, no packing
+            grouped = bf16_linear(attn_out, oa_w.cuda())
+        elif oa_w.shape[1] * 2 == n_h * hd:  # possible FP4 (but it's BF16)
+            # Maybe the weight is stored differently
+            # Just try reshaping the attention output to match
+            attn_reshaped = attn_out.reshape(T, -1)[:, :oa_w.shape[1]]
+            grouped = bf16_linear(attn_reshaped, oa_w.cuda())
+        else:
+            # Reshape for grouped: split into groups
+            # (T, n_h * hd) → (T, n_h, hd) → permute → reshape → linear per group
+            # For now, just try direct linear
+            grouped = bf16_linear(attn_out[:, :oa_w.shape[1]], oa_w.cuda())
+    except RuntimeError:
+        # Fallback: pad or truncate
+        if attn_out.shape[-1] < oa_w.shape[1]:
+            padded = torch.nn.functional.pad(attn_out, (0, oa_w.shape[1] - attn_out.shape[-1]))
+            grouped = bf16_linear(padded, oa_w.cuda())
+        else:
+            grouped = bf16_linear(attn_out[:, :oa_w.shape[1]], oa_w.cuda())
+    
     attn_proj = nvfp4_linear(grouped, ob_w, ob_s, ob_s2)  # (1, H)
     
     # ---- Residual ----