#!/usr/bin/env python3 """Test fused SwiGLU NVFP4 GEMM kernel compilation and correctness. Validates P0/P1 from PERFORMANCE_AUDIT.md: - Fused SwiGLU kernel compiles via cute.compile - Output cosine similarity vs unfused path >= 0.9995 - Tests both multi-expert (MoE) and single-expert (SharedExpert) modes """ import torch import sys def test_fused_swiglu_compilation(): """Test that the fused SwiGLU kernel compiles and runs.""" from dsv4.ops.gemm_runner import ( warmup_fused_swiglu_compilation, warmup_compilation, run_nvfp4_grouped_gemm, run_fused_swiglu_grouped_gemm, ) from dsv4.ops.quantize import quantize_to_nvfp4, SF_VEC_SIZE from dsv4.ops.layouts import make_b_k_major, interleave_l1_weights device = "cuda:0" # Production MoE shapes (DeepSeek-V4 Pro L1 GEMM) # L1: K=7168, N=6144 (gate+up combined) → K_packed=3584, N_packed=3072 K_packed = 3584 N_packed = 3072 num_experts = 4 # Small for testing, but >1 for MoE path swiglu_limit = 10.0 print(f"Testing fused SwiGLU kernel compilation...") print(f" K_packed={K_packed}, N_packed={N_packed}, E={num_experts}, limit={swiglu_limit}") # Warmup standard GEMM first print(" Warming up standard GEMM...") warmup_compilation(num_experts, K_packed, N_packed, device) # Warmup fused GEMM print(" Warming up fused SwiGLU GEMM...") try: warmup_fused_swiglu_compilation( num_experts, K_packed, N_packed, device, swiglu_limit=swiglu_limit, ) print(" ✅ Fused SwiGLU kernel compiled successfully!") except TypeError as e: print(f" ❌ Fused SwiGLU compilation FAILED with TypeError: {e}") print(f" This is the arg-binding bug from the previous session.") raise except Exception as e: print(f" ❌ Fused SwiGLU compilation FAILED: {type(e).__name__}: {e}") raise # Now test correctness: run both fused and unfused, compare print("\n Testing fused vs unfused output correctness...") tokens = 6 # top-k=6 K = K_packed * 2 # 7168 N = N_packed * 2 # 6144 # Create random input x_bf16 = torch.randn(tokens, K, dtype=torch.bfloat16, device=device) * 0.5 # Create random weight (same for both paths) w_bf16 = torch.randn(num_experts, K, N, dtype=torch.bfloat16, device=device) * 0.1 # Quantize activation x_fp4, x_sf, x_gs = quantize_to_nvfp4(x_bf16) # Quantize weight (interleaved for L1 gate+up) w_bf16_t = w_bf16.permute(0, 2, 1).contiguous() # (E, N, K) for make_b_k_major w_fp4, w_sf, w_gs = quantize_to_nvfp4(w_bf16_t) # w_fp4: (E, N_packed, K_packed) — interleave along N for gate/up pairing if w_fp4.dtype == torch.uint8: w_fp4 = w_fp4.view(torch.float4_e2m1fn_x2) w_fp4_il = interleave_l1_weights(w_fp4) # (E, N_packed, K_packed) interleaved mat_b = make_b_k_major(w_fp4_il) # Expert offsets (all tokens go to expert 0 for simplicity) expert_offsets = torch.tensor([0, tokens], dtype=torch.int32, device=device) padded_offsets = torch.tensor([128], dtype=torch.int32, device=device) # padded to 128 # Pad activation to 128 rows x_padded = torch.zeros(128, K_packed, dtype=x_fp4.dtype, device=device) x_padded[:tokens] = x_fp4 # Assemble scales (simplified — just pad + swizzle) from dsv4.ops.layouts import pad_and_swizzle_single, ceil_div as cutedsl_ceil_div K_sf = cutedsl_ceil_div(K, 16) padded_cols = cutedsl_ceil_div(K_sf, 4) * 4 scale_a_buf = torch.zeros(128, padded_cols, dtype=torch.float16, device=device).to(torch.float8_e4m3fn) scale_a_buf[:tokens, :x_sf.shape[1]] = x_sf scale_a = pad_and_swizzle_single(scale_a_buf).reshape(128, padded_cols) from dsv4.ops.layouts import assemble_scales_3d_side scale_b = assemble_scales_3d_side(w_sf) global_scale_a = torch.full((num_experts,), x_gs, dtype=torch.float32, device=device) global_scale_b = torch.tensor(w_gs, dtype=torch.float32, device=device) # Run UNFUSED path print(" Running unfused GEMM...") l1_unfused = run_nvfp4_grouped_gemm( mat_a=x_padded, mat_b=mat_b, scale_a=scale_a, scale_b=scale_b, expert_offsets=padded_offsets, global_scale_a=global_scale_a, global_scale_b=global_scale_b, )[:tokens] # (6, 6144) BF16 # Manual SwiGLU on unfused output intermediate = N // 2 # 3072 l1_deil = interleave_l1_weights(l1_unfused.unsqueeze(0).contiguous())[0] gate = l1_deil[:, :intermediate] up = l1_deil[:, intermediate:] gate_silu = torch.nn.functional.silu(gate) gate_silu = gate_silu.clamp(max=swiglu_limit) up = up.clamp(min=-swiglu_limit, max=swiglu_limit) activated_unfused = gate_silu * up # Run FUSED path print(" Running fused SwiGLU GEMM...") try: l1_fused = run_fused_swiglu_grouped_gemm( mat_a=x_padded, mat_b=mat_b, scale_a=scale_a, scale_b=scale_b, expert_offsets=padded_offsets, global_scale_a=global_scale_a, global_scale_b=global_scale_b, swiglu_limit=swiglu_limit, )[:tokens] # (6, 3072) BF16 — SwiGLU already applied print(" ✅ Fused SwiGLU GEMM ran successfully!") except Exception as e: print(f" ❌ Fused SwiGLU GEMM FAILED: {type(e).__name__}: {e}") raise # Compare # The fused kernel outputs only the silu(gate)*up result (N/2 = 3072) # The unfused path's activated_unfused is the same computation in Python cos = torch.nn.functional.cosine_similarity( l1_fused.flatten().float(), activated_unfused.flatten().float(), dim=0 ).item() max_diff = (l1_fused.float() - activated_unfused.float()).abs().max().item() print(f"\n Fused vs Unfused SwiGLU output:") print(f" Cosine similarity: {cos:.6f}") print(f" Max abs diff: {max_diff:.6f}") print(f" |fused|: {l1_fused.abs().max().item():.4f}") print(f" |unfused|: {activated_unfused.abs().max().item():.4f}") if cos >= 0.9995: print(f" ✅ PASS: cosine >= 0.9995") else: print(f" ❌ FAIL: cosine < 0.9995") # Test single-expert mode (for SharedExpert P1) print("\n--- Testing single-expert (SharedExpert) mode ---") warmup_fused_swiglu_compilation( 1, K_packed, N_packed, device, swiglu_limit=swiglu_limit ) print(" ✅ Single-expert fused SwiGLU compiled!") return cos >= 0.9995 if __name__ == "__main__": torch.manual_seed(42) success = test_fused_swiglu_compilation() sys.exit(0 if success else 1)