#!/usr/bin/env python3 """ Test A: Compare moe_pipeline output vs CuTeDSLMoERunner output. Uses the same weights and inputs. If they differ, the runner is broken. Runs on the B200 host (not inside Docker): source /root/nvfp4-megamoe-kernel/tests/.venv/bin/activate python3 tests/test_runner_vs_pipeline.py """ import os, sys, json, torch from safetensors import safe_open REPO_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) sys.path.insert(0, REPO_ROOT) from cutedsl.moe_pipeline import run_nvfp4_moe from vllm.nvfp4_cutedsl import CuTeDSLMoERunner from cutedsl.bridge import quantize_to_nvfp4, quantize_weight_to_nvfp4, make_b_k_major, assemble_scales_3d_side, compute_expert_offsets MODEL_DIR = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4" DEVICE = "cuda" LAYER_IDX = 0 E2M1_LUT = torch.tensor([0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0, -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0], dtype=torch.float32) def find_shards(model_dir): index_path = os.path.join(model_dir, "model.safetensors.index.json") key_to_shard = {} if os.path.exists(index_path): with open(index_path) as f: index = json.load(f) for key, shard in index["weight_map"].items(): key_to_shard[key] = os.path.join(model_dir, shard) return key_to_shard def load_layer_tensors(model_dir, layer_idx): key_to_shard = find_shards(model_dir) layer_prefix = f"layers.{layer_idx}." tensors = {} for key, shard in key_to_shard.items(): norm_key = key.removeprefix("model.") if not norm_key.startswith(layer_prefix): continue with safe_open(shard, framework="pt") as f: if key in f.keys(): tensors[norm_key] = f.get_tensor(key) return tensors def dequantize_nvfp4_weight(packed_uint8, scale_e4m3, global_scale): device = packed_uint8.device lut = E2M1_LUT.to(device) lower = lut[(packed_uint8 & 0x0F).long()] upper = lut[((packed_uint8 >> 4) & 0x0F).long()] out_features = packed_uint8.shape[0] in_features = packed_uint8.shape[1] * 2 unpacked = torch.empty(out_features, in_features, dtype=torch.float32, device=device) unpacked[:, 0::2] = lower unpacked[:, 1::2] = upper block_scale = scale_e4m3.float() block_expanded = block_scale.repeat_interleave(16, dim=1)[:, :in_features] return (unpacked * block_expanded * global_scale).to(torch.bfloat16) def prepare_direct_weights(nvfp4_tensors, layer_idx, expert_indices, intermediate_size): """Direct view-cast path (same as layertest).""" l1_fp4, l1_sf, l1_gs = [], [], [] l2_fp4, l2_sf, l2_gs = [], [], [] for e in expert_indices: gate_w = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.gate_proj.weight"].to(DEVICE) up_w = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.up_proj.weight"].to(DEVICE) gate_sf = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.gate_proj.weight_scale"].to(DEVICE) up_sf = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.up_proj.weight_scale"].to(DEVICE) gate_gs = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.gate_proj.weight_scale_2"].item() up_gs = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.up_proj.weight_scale_2"].item() fused_w = torch.cat([gate_w, up_w], dim=0) fused_w_fp4 = fused_w.view(torch.float4_e2m1fn_x2).permute(1, 0).contiguous() fused_sf = torch.cat([gate_sf, up_sf], dim=0).permute(1, 0).contiguous() max_gs = max(gate_gs, up_gs) if gate_gs != up_gs: f32 = fused_sf.float() f32[:, :intermediate_size] *= (gate_gs / max_gs) f32[:, intermediate_size:] *= (up_gs / max_gs) fused_sf = f32.to(torch.float8_e4m3fn) l1_fp4.append(fused_w_fp4) l1_sf.append(fused_sf) l1_gs.append(max_gs) down_w = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.down_proj.weight"].to(DEVICE) down_sf = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.down_proj.weight_scale"].to(DEVICE) down_gs = nvfp4_tensors[f"layers.{layer_idx}.mlp.experts.{e}.down_proj.weight_scale_2"].item() l2_fp4.append(down_w.view(torch.float4_e2m1fn_x2).permute(1, 0).contiguous()) l2_sf.append(down_sf.permute(1, 0).contiguous()) l2_gs.append(down_gs) return {'l1_fp4': l1_fp4, 'l1_sf': l1_sf, 'l1_gs': l1_gs, 'l2_fp4': l2_fp4, 'l2_sf': l2_sf, 'l2_gs': l2_gs} def main(): torch.manual_seed(42) expert_indices = [0, 1, 2] num_experts = len(expert_indices) hidden_size = 7168 intermediate_size = 3072 top_k = 2 num_tokens = 4 print("=" * 70) print(" Loading checkpoint") print("=" * 70) nvfp4_tensors = load_layer_tensors(MODEL_DIR, LAYER_IDX) print(f" {len(nvfp4_tensors)} tensors loaded") weights = prepare_direct_weights(nvfp4_tensors, LAYER_IDX, expert_indices, intermediate_size) hidden_states = torch.randn(num_tokens, hidden_size, dtype=torch.bfloat16, device=DEVICE) * 2.0 expert_ids = torch.tensor([[0, 1]] * num_tokens, dtype=torch.int32, device=DEVICE) expert_weights = torch.tensor([[0.6, 0.4]] * num_tokens, dtype=torch.float32, device=DEVICE) # ── Path 1: moe_pipeline (reference, uses quantize_to_nvfp4) ── print("\n Running moe_pipeline (dynamic gs)...") pipeline_out = run_nvfp4_moe( hidden_states.clone(), expert_ids.clone(), expert_weights.clone(), weights, expert_indices, ) print(f" Pipeline: amax={pipeline_out.abs().max():.4f}, mean={pipeline_out.float().mean():.6f}") # ── Path 2: CuTeDSLMoERunner with checkpoint input_scale (what vLLM uses) ── print("\n Running CuTeDSLMoERunner (checkpoint gs)...") runner = CuTeDSLMoERunner(num_experts, hidden_size, intermediate_size, device=DEVICE) runner.prepare_weights_direct( [w.clone() for w in weights['l1_fp4']], [w.clone() for w in weights['l1_sf']], list(weights['l1_gs']), [w.clone() for w in weights['l2_fp4']], [w.clone() for w in weights['l2_sf']], list(weights['l2_gs']), ) # Set checkpoint input_scale (what vLLM does in finalize_weights) igs = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.0.gate_proj.input_scale"].item() runner._l1_activation_global_scale = igs runner._l2_activation_global_scale = igs print(f" Checkpoint input_scale: {igs:.10f}") # Build topk_weights and topk_ids in the format the runner expects # runner.run expects topk_ids as expert indices (0-based within our expert set) topk_weights = expert_weights topk_ids = expert_ids runner_out = runner.run(hidden_states.clone(), topk_weights, topk_ids) print(f" Runner (ckpt gs): amax={runner_out.abs().max():.4f}, mean={runner_out.float().mean():.6f}") cos_ckpt = torch.nn.functional.cosine_similarity( runner_out.flatten().unsqueeze(0).float(), pipeline_out.flatten().unsqueeze(0).float(), ).item() print(f" Cosine vs pipeline: {cos_ckpt:.6f}") # ── Path 3: CuTeDSLMoERunner with dynamic gs ── print("\n Running CuTeDSLMoERunner (dynamic gs)...") # We can't use quantize_to_nvfp4 in the runner (cudagraph), but we can # compute the gs from the input and set it before calling run x_igs = (hidden_states.abs().max().item()) / (6.0 * 448.0) runner2 = CuTeDSLMoERunner(num_experts, hidden_size, intermediate_size, device=DEVICE) runner2.prepare_weights_direct( [w.clone() for w in weights['l1_fp4']], [w.clone() for w in weights['l1_sf']], list(weights['l1_gs']), [w.clone() for w in weights['l2_fp4']], [w.clone() for w in weights['l2_sf']], list(weights['l2_gs']), ) runner2._l1_activation_global_scale = x_igs runner2._l2_activation_global_scale = x_igs print(f" Dynamic gs (from input amax): {x_igs:.10f}") runner2_out = runner2.run(hidden_states.clone(), topk_weights, topk_ids) print(f" Runner (dynamic gs): amax={runner2_out.abs().max():.4f}, mean={runner2_out.float().mean():.6f}") cos_dyn = torch.nn.functional.cosine_similarity( runner2_out.flatten().unsqueeze(0).float(), pipeline_out.flatten().unsqueeze(0).float(), ).item() print(f" Cosine vs pipeline: {cos_dyn:.6f}") # ── Summary ── print(f"\n{'=' * 70}") print(f" RESULTS") print(f"{'=' * 70}") print(f" Runner with checkpoint gs vs pipeline: {cos_ckpt:.6f}") print(f" Runner with dynamic gs vs pipeline: {cos_dyn:.6f}") if cos_dyn > 0.95: print(f" ✅ Dynamic gs fixes the problem — gs is the only bug") elif cos_dyn < 0.5 and cos_ckpt < 0.5: print(f" ❌ Both runner paths are broken — scale assembly is also wrong") else: print(f" ⚠️ Partial match — multiple issues") if __name__ == "__main__": main()