submission 407575

  ⋯ 5 unchanged lines
  def _fused_prologue_kernel(
      X_ptr, M_ptr, L_ptr, R_ptr, OG_ptr,
      W_norm_ptr, B_norm_ptr,
-     W_L_ptr, W_R_ptr, W_LG_ptr, W_RG_ptr, W_OG_ptr,
+     W_PACKED_ptr,
      stride_xb, stride_xi, stride_xj, stride_xc,
      stride_mb, stride_mi, stride_mj,
+     stride_l_b, stride_l_d, stride_l_i, stride_l_j,
+     stride_r_b, stride_r_d, stride_r_i, stride_r_j,
+     stride_og_b, stride_og_i, stride_og_j, stride_og_d,
      B, N, C, D: tl.constexpr,
      BLOCK_SIZE_N: tl.constexpr,
      BLOCK_SIZE_C: tl.constexpr
  ⋯ 5 unchanged lines
      offsets_j = pid_j_start + tl.arange(0, BLOCK_SIZE_N)
      mask_j = offsets_j < N
   
-     # 1. LayerNorm statistics
+     # 1. LayerNorm statistics (Online algorithm)
      acc_sum = tl.zeros([BLOCK_SIZE_N], dtype=tl.float32)
      acc_sum_sq = tl.zeros([BLOCK_SIZE_N], dtype=tl.float32)
      
  ⋯ 9 unchanged lines
      var = (acc_sum_sq / C) - (mean * mean)
      rstd = 1.0 / tl.sqrt(var + 1e-5)
   
-     # 2. Projections
+     # 2. Projections with Packed Load, Separate Compute
      acc_l = tl.zeros([BLOCK_SIZE_N, D], dtype=tl.float32)
      acc_lg = tl.zeros([BLOCK_SIZE_N, D], dtype=tl.float32)
      acc_r = tl.zeros([BLOCK_SIZE_N, D], dtype=tl.float32)
  ⋯ 10 unchanged lines
          x_chunk = tl.load(x_ptr, mask=(mask_j[:, None] & c_mask[None, :]), other=0.0).to(tl.float32)
          w_n = tl.load(W_norm_ptr + cols, mask=c_mask, other=0.0)
          b_n = tl.load(B_norm_ptr + cols, mask=c_mask, other=0.0)
-         x_n = (x_chunk - mean[:, None]) * rstd[:, None] * w_n[None, :] + b_n[None, :]
-         x_n = x_n.to(tl.float16)
+         x_n = ((x_chunk - mean[:, None]) * rstd[:, None] * w_n[None, :] + b_n[None, :]).to(tl.float16)
   
-         # Load weights and perform dots
-         acc_l += tl.dot(x_n, tl.load(W_L_ptr + cols[:, None] * D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
-         acc_lg += tl.dot(x_n, tl.load(W_LG_ptr + cols[:, None] * D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
-         acc_r += tl.dot(x_n, tl.load(W_R_ptr + cols[:, None] * D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
-         acc_rg += tl.dot(x_n, tl.load(W_RG_ptr + cols[:, None] * D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
-         acc_og += tl.dot(x_n, tl.load(W_OG_ptr + cols[:, None] * D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
+         # Load all 5 weights in one contiguous block [C, 5*D]
+         # We use separate tl.dot to avoid register slicing errors
+         w_base = W_PACKED_ptr + cols[:, None] * (5 * D)
+         acc_l  += tl.dot(x_n, tl.load(w_base + 0*D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
+         acc_lg += tl.dot(x_n, tl.load(w_base + 1*D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
+         acc_r  += tl.dot(x_n, tl.load(w_base + 2*D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
+         acc_rg += tl.dot(x_n, tl.load(w_base + 3*D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
+         acc_og += tl.dot(x_n, tl.load(w_base + 4*D + off_d[None, :], mask=c_mask[:, None], other=0.0).to(tl.float16))
   
+     # 3. Gating and Masking
      m_ptr = M_ptr + pid_b * stride_mb + pid_i * stride_mi + offsets_j
      mask_val = tl.load(m_ptr, mask=mask_j, other=0.0).to(tl.float32)
   
  ⋯ 1 unchanged lines
      r_final = acc_r * tl.sigmoid(acc_rg) * mask_val[:, None]
      og_final = tl.sigmoid(acc_og)
   
-     # Store L, R in [B, D, N, N] for BMM, OG in [B, N, N, D]
+     # 4. Stores for BMM [B, D, N, N]
      idx_nn = pid_i * N + offsets_j
      off_l_r = pid_b * D * N * N + off_d[None, :] * N * N + idx_nn[:, None]
      tl.store(L_ptr + off_l_r, l_final.to(tl.float16), mask=(mask_j[:, None] & (off_d[None, :] < D)))
      tl.store(R_ptr + off_l_r, r_final.to(tl.float16), mask=(mask_j[:, None] & (off_d[None, :] < D)))
      
-     off_og = pid_b * N * N * D + idx_nn[:, None] * D + off_d[None, :]
+     off_og = pid_b * stride_og_b + pid_i * stride_og_i + offsets_j[:, None] * stride_og_j + off_d[None, :]
      tl.store(OG_ptr + off_og, og_final.to(tl.float16), mask=(mask_j[:, None] & (off_d[None, :] < D)))
   
  @triton.jit
- def _fused_epilogue_kernel(
+ def _super_epilogue_kernel(
      BMM_OUT_ptr, OG_ptr, OUT_ptr,
-     W_TN_ptr, B_TN_ptr,
+     W_TN_ptr, B_TN_ptr, W_TO_ptr,
      stride_bmm_b, stride_bmm_d, stride_bmm_i, stride_bmm_j,
      stride_og_b, stride_og_i, stride_og_j, stride_og_d,
-     stride_out_b, stride_out_i, stride_out_j, stride_out_d,
-     B, N, D: tl.constexpr,
-     BLOCK_SIZE_N: tl.constexpr
+     stride_out_b, stride_out_i, stride_out_j, stride_out_c,
+     B, N, C, D: tl.constexpr,
+     BLOCK_SIZE_N: tl.constexpr,
+     BLOCK_SIZE_C: tl.constexpr
  ):
      pid_b = tl.program_id(0)
      pid_i = tl.program_id(1)
  ⋯ 3 unchanged lines
      mask_j = offsets_j < N
      off_d = tl.arange(0, D)
   
-     # Load BMM output [BLOCK_SIZE_N, D] from [B, D, N, N]
      bmm_ptr = BMM_OUT_ptr + pid_b * stride_bmm_b + off_d[None, :] * stride_bmm_d + pid_i * stride_bmm_i + offsets_j[:, None] * stride_bmm_j
      val = tl.load(bmm_ptr, mask=(mask_j[:, None] & (off_d[None, :] < D)), other=0.0).to(tl.float32)
      
-     # Load OG [BLOCK_SIZE_N, D] from [B, N, N, D]
      og_ptr = OG_ptr + pid_b * stride_og_b + pid_i * stride_og_i + offsets_j[:, None] * stride_og_j + off_d[None, :]
      og = tl.load(og_ptr, mask=(mask_j[:, None] & (off_d[None, :] < D)), other=0.0).to(tl.float32)
   
-     # LayerNorm over D
      mean = tl.sum(val, axis=1) / D
      var = (tl.sum(val * val, axis=1) / D) - (mean * mean)
      rstd = 1.0 / tl.sqrt(var + 1e-5)
  ⋯ 1 unchanged lines
      w_tn = tl.load(W_TN_ptr + off_d)
      b_tn = tl.load(B_TN_ptr + off_d)
      val = (val - mean[:, None]) * rstd[:, None] * w_tn[None, :] + b_tn[None, :]
-     val = val * og
+     val = (val * og).to(tl.float16)
   
-     # Store in [B, N, N, D] format for final matmul
-     out_ptr = OUT_ptr + pid_b * stride_out_b + pid_i * stride_out_i + offsets_j[:, None] * stride_out_j + off_d[None, :]
-     tl.store(out_ptr, val.to(tl.float16), mask=(mask_j[:, None] & (off_d[None, :] < D)))
+     for c_offset in range(0, C, BLOCK_SIZE_C):
+         off_c = c_offset + tl.arange(0, BLOCK_SIZE_C)
+         c_mask = off_c < C
+         w_to = tl.load(W_TO_ptr + off_c[None, :] * D + off_d[:, None], mask=c_mask[None, :], other=0.0).to(tl.float16)
+         out_chunk = tl.dot(val, w_to)
+         
+         out_ptr = OUT_ptr + pid_b * stride_out_b + pid_i * stride_out_i + offsets_j[:, None] * stride_out_j + off_c[None, :]
+         tl.store(out_ptr, out_chunk.to(tl.float32), mask=(mask_j[:, None] & c_mask[None, :]))
   
  def custom_kernel(data):
      x, mask, weights, config = data
  ⋯ 2 unchanged lines
      device = x.device
      w_fp16 = {k: v.to(torch.float16) for k, v in weights.items()}
   
+     # Pack all 5 weights: [C, 5*D]
+     w_packed = torch.cat([
+         w_fp16["left_proj.weight"],
+         w_fp16["left_gate.weight"],
+         w_fp16["right_proj.weight"],
+         w_fp16["right_gate.weight"],
+         w_fp16["out_gate.weight"]
+     ], dim=0).t().contiguous()
+  
      L = torch.empty((B, D, N, N), device=device, dtype=torch.float16)
      R = torch.empty((B, D, N, N), device=device, dtype=torch.float16)
      OG = torch.empty((B, N, N, D), device=device, dtype=torch.float16)
   
-     # Prologue: LN + 5 Projections + Gating. Reduced BLOCK_SIZE_C and num_stages=1 to fit SRAM.
      grid_pre = (B, N, (N + 32 - 1) // 32)
      _fused_prologue_kernel[grid_pre](
          x, mask, L, R, OG,
          w_fp16["norm.weight"], w_fp16["norm.bias"],
-         w_fp16["left_proj.weight"].t().contiguous(), w_fp16["right_proj.weight"].t().contiguous(),
-         w_fp16["left_gate.weight"].t().contiguous(), w_fp16["right_gate.weight"].t().contiguous(),
-         w_fp16["out_gate.weight"].t().contiguous(),
+         w_packed,
          x.stride(0), x.stride(1), x.stride(2), x.stride(3),
          mask.stride(0), mask.stride(1), mask.stride(2),
-         B, N, C, D, BLOCK_SIZE_N=32, BLOCK_SIZE_C=32, num_stages=1
+         L.stride(0), L.stride(1), L.stride(2), L.stride(3),
+         R.stride(0), R.stride(1), R.stride(2), R.stride(3),
+         OG.stride(0), OG.stride(1), OG.stride(2), OG.stride(3),
+         B, N, C, D, BLOCK_SIZE_N=32, BLOCK_SIZE_C=32
      )
   
-     # Contraction: [B*D, N, N] @ [B*D, N, N].T -> [B, D, N, N]
      bmm_out = torch.bmm(L.view(B * D, N, N), R.view(B * D, N, N).transpose(-1, -2)).view(B, D, N, N)
      
-     # Epilogue Part 1: LN + Gating. Fused to handle the [B, D, N, N] -> [B, N, N, D] layout change.
-     epi_inter = torch.empty((B, N, N, D), device=device, dtype=torch.float16)
+     output = torch.empty((B, N, N, C), device=device, dtype=torch.float32)
      grid_epi = (B, N, (N + 32 - 1) // 32)
-     _fused_epilogue_kernel[grid_epi](
-         bmm_out, OG, epi_inter,
-         w_fp16["to_out_norm.weight"], w_fp16["to_out_norm.bias"],
+     _super_epilogue_kernel[grid_epi](
+         bmm_out, OG, output,
+         w_fp16["to_out_norm.weight"], w_fp16["to_out_norm.bias"], w_fp16["to_out.weight"],
          bmm_out.stride(0), bmm_out.stride(1), bmm_out.stride(2), bmm_out.stride(3),
          OG.stride(0), OG.stride(1), OG.stride(2), OG.stride(3),
-         epi_inter.stride(0), epi_inter.stride(1), epi_inter.stride(2), epi_inter.stride(3),
-         B, N, D, BLOCK_SIZE_N=32
+         output.stride(0), output.stride(1), output.stride(2), output.stride(3),
+         B, N, C, D, BLOCK_SIZE_N=32, BLOCK_SIZE_C=64
      )
   
-     # Epilogue Part 2: Final Projection to C using cuBLAS.
-     return (epi_inter @ w_fp16["to_out.weight"].t()).to(torch.float32)
+     return output