Full pipeline test: runner vs BF16 reference

This commit is contained in:
2026-05-17 21:29:16 +00:00
parent 2796bd81e8
commit 72628fb689

View File

@@ -1,6 +1,4 @@
"""Debug test: Replicate runner logic step by step in Python.
Compare against BF16 reference to isolate where tokens get dropped.
"""
"""Full pipeline test: Fixed runner vs BF16 reference."""
import torch
import torch.nn.functional as F
import sys, os, glob
@@ -45,8 +43,9 @@ def main():
torch.cuda.set_device(0)
torch.manual_seed(42)
print("=== Runner Logic Debug ===")
print("=== Full Pipeline Test (Fixed Runner) ===")
nvfp4_tensors = load_layer_tensors(MODEL_PATH, LAYER_IDX)
expert_indices = list(range(NUM_EXPERTS))
hidden_states = torch.randn(NUM_TOKENS, HIDDEN_SIZE, dtype=torch.bfloat16, device=DEVICE) * 2.0
topk_ids = torch.zeros(NUM_TOKENS, TOP_K, dtype=torch.int64, device=DEVICE)
@@ -54,89 +53,48 @@ def main():
topk_ids[i] = torch.randperm(NUM_EXPERTS)[:TOP_K]
topk_weights = torch.ones(NUM_TOKENS, TOP_K, dtype=torch.float32, device=DEVICE) / TOP_K
# Step 1: Global→local remap (same as runner)
experts_start_idx = 0
local_ids = topk_ids - experts_start_idx
local_mask = (local_ids >= 0) & (local_ids < NUM_EXPERTS)
safe_ids = local_ids.clamp(0, NUM_EXPERTS - 1)
safe_weights = topk_weights * local_mask.float()
# BF16 reference
ref_out = torch.zeros(NUM_TOKENS, HIDDEN_SIZE, dtype=torch.bfloat16, device=DEVICE)
for i, e in enumerate(expert_indices):
dk = f"layers.{LAYER_IDX}.mlp.experts.{e}.down_proj.weight"
gk = f"layers.{LAYER_IDX}.mlp.experts.{e}.gate_proj.weight"
uk = f"layers.{LAYER_IDX}.mlp.experts.{e}.up_proj.weight"
if dk not in nvfp4_tensors:
continue
gate_bf16 = dequantize_nvfp4_weight(
nvfp4_tensors[gk].to(DEVICE),
nvfp4_tensors[gk.replace('.weight', '.weight_scale')].to(DEVICE),
nvfp4_tensors[gk.replace('.weight', '.weight_scale_2')].item())
up_bf16 = dequantize_nvfp4_weight(
nvfp4_tensors[uk].to(DEVICE),
nvfp4_tensors[uk.replace('.weight', '.weight_scale')].to(DEVICE),
nvfp4_tensors[uk.replace('.weight', '.weight_scale_2')].item())
down_bf16 = dequantize_nvfp4_weight(
nvfp4_tensors[dk].to(DEVICE),
nvfp4_tensors[dk.replace('.weight', '.weight_scale')].to(DEVICE),
nvfp4_tensors[dk.replace('.weight', '.weight_scale_2')].item())
for t in range(NUM_TOKENS):
for k in range(TOP_K):
if topk_ids[t, k].item() != i:
continue
w = topk_weights[t, k].item()
x = hidden_states[t]
gate = x @ gate_bf16.T
up = x @ up_bf16.T
gate_silu = F.silu(gate).clamp(max=SWIGLU_LIMIT)
up = up.clamp(min=-SWIGLU_LIMIT, max=SWIGLU_LIMIT)
act = gate_silu * up
ref_out[t] += w * (act @ down_bf16.T)
print(f"topk_ids:\n{topk_ids}")
print(f"local_ids:\n{local_ids}")
print(f"local_mask:\n{local_mask}")
print(f"safe_weights (should all be 0.1667):\n{safe_weights}")
print(f"BF16 ref: amax={ref_out.amax().item():.4f}")
# Step 2: Sort by expert
flat_ids = safe_ids.reshape(-1)
flat_weights = safe_weights.reshape(-1)
num_slots = NUM_TOKENS * TOP_K
token_indices = torch.arange(NUM_TOKENS, device=DEVICE).unsqueeze(1).expand(-1, TOP_K).reshape(-1)
# CuTeDSL runner
from vllm.nvfp4_cutedsl import CuTeDSLMoERunner
from cutedsl.bridge import assemble_scales_3d_side, make_b_k_major
sort_idx = flat_ids.argsort(stable=True)
sorted_ids = flat_ids[sort_idx]
sorted_weights = flat_weights[sort_idx]
sorted_token_ids = token_indices[sort_idx]
print(f"\nsorted_ids: {sorted_ids.tolist()}")
print(f"sorted_token_ids: {sorted_token_ids.tolist()}")
print(f"sorted_weights: {sorted_weights.tolist()}")
# Step 3: Expert offsets
expert_id_range = torch.arange(NUM_EXPERTS, device=DEVICE)
tokens_per_expert = (sorted_ids.unsqueeze(1) == expert_id_range.unsqueeze(0)).sum(dim=0).int()
expert_offsets = torch.zeros(NUM_EXPERTS + 1, dtype=torch.int32, device=DEVICE)
expert_offsets[1:] = tokens_per_expert.cumsum(0)
print(f"\ntokens_per_expert: {tokens_per_expert.tolist()}")
print(f"expert_offsets: {expert_offsets.tolist()}")
# Step 4: Padded offsets
padded_tokens_per_expert = ((tokens_per_expert + 127) // 128) * 128
padded_expert_offsets = torch.zeros(NUM_EXPERTS + 1, dtype=torch.int32, device=DEVICE)
padded_expert_offsets[1:] = padded_tokens_per_expert.cumsum(0)
total_padded = padded_expert_offsets[NUM_EXPERTS].item()
print(f"padded_tokens_per_expert: {padded_tokens_per_expert.tolist()}")
print(f"padded_expert_offsets: {padded_expert_offsets.tolist()}")
print(f"total_padded: {total_padded}")
# Step 5: Scatter into padded layout (runner's searchsorted approach)
row_indices = torch.arange(num_slots, device=DEVICE)
expert_assign = torch.searchsorted(expert_offsets[1:], row_indices, right=True).clamp(max=NUM_EXPERTS - 1)
local_row = row_indices - expert_offsets[expert_assign]
padded_dst = padded_expert_offsets[expert_assign] + local_row
print(f"\nexpert_assign: {expert_assign.tolist()}")
print(f"local_row: {local_row.tolist()}")
print(f"padded_dst: {padded_dst.tolist()}")
# Verify: expert_assign should match sorted_ids
match = (expert_assign == sorted_ids).all().item()
print(f"expert_assign == sorted_ids: {match}")
if not match:
mismatches = (expert_assign != sorted_ids).nonzero().squeeze()
print(f" Mismatch at rows: {mismatches.tolist()}")
print(f" expert_assign[mismatch]: {expert_assign[mismatches].tolist()}")
print(f" sorted_ids[mismatch]: {sorted_ids[mismatches].tolist()}")
# Step 6: Scatter hidden states
slot_hidden = hidden_states[sorted_token_ids]
padded_hidden = torch.zeros(total_padded, HIDDEN_SIZE, dtype=torch.bfloat16, device=DEVICE)
padded_hidden[padded_dst] = slot_hidden
# Verify: padded_hidden[padded_dst] should match slot_hidden
verify = (padded_hidden[padded_dst] == slot_hidden).all().item()
print(f"\npadded_hidden scatter correct: {verify}")
# Step 7: Now run L1 GEMM using bridge (direct call, not runner)
from cutedsl.bridge import (
quantize_to_nvfp4, run_nvfp4_grouped_gemm,
assemble_scales_3d_side, make_b_k_major,
)
# Prepare weights (same as runner's _ensure_stacked)
expert_indices = list(range(NUM_EXPERTS))
l1_fp4, l1_sf, l1_gs_list = [], [], []
l1_fp4, l1_sf, l1_gs = [], [], []
l2_fp4, l2_sf, l2_gs = [], [], []
for e in expert_indices:
gw = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.gate_proj.weight"].to(DEVICE)
uw = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.up_proj.weight"].to(DEVICE)
@@ -148,106 +106,54 @@ def main():
fsf = torch.cat([gsf, usf], dim=0).permute(1,0).contiguous()
mgs = max(ggs, ugs)
if ggs != ugs:
fsf32 = fsf.float()
fsf32[:, :INTERMEDIATE_SIZE] *= (ggs / mgs)
fsf32[:, INTERMEDIATE_SIZE:] *= (ugs / mgs)
fsf = fsf32.to(torch.float8_e4m3fn)
l1_fp4.append(fw); l1_sf.append(fsf); l1_gs_list.append(mgs)
l1_mat_b = torch.stack(l1_fp4)
l1_mat_b = make_b_k_major(l1_mat_b)
l1_scale_b = assemble_scales_3d_side(l1_sf)
l1_gsb = torch.tensor(l1_gs_list, dtype=torch.float32, device=DEVICE)
# Quantize activation (dynamic gs, not warmup)
# KEY FIX: quantize slot_hidden (sorted tokens), NOT padded_hidden.
# padded_hidden has zeros in padding rows; quantizing it gives wrong x_sf layout.
print("\n--- L1 GEMM (dynamic gs) ---")
slot_x_fp4, slot_x_sf, l1_gs = quantize_to_nvfp4(slot_hidden)
print(f" L1 gs (dynamic): {l1_gs:.6f}")
# Scatter x_fp4 into padded layout (use uint8 for scatter, then view as float4)
padded_x_fp4_uint8 = torch.zeros(total_padded, HIDDEN_SIZE // 2, dtype=torch.uint8, device=DEVICE)
padded_x_fp4_uint8[padded_dst] = slot_x_fp4.view(torch.uint8)
padded_x_fp4 = padded_x_fp4_uint8.view(torch.float4_e2m1fn_x2)
# For scale_a, we need to use the runner's assembly approach.
# Use the same _assemble_scales_cudagraph_safe function
from vllm.nvfp4_cutedsl import CuTeDSLMoERunner
runner = CuTeDSLMoERunner(
num_experts=NUM_EXPERTS, hidden_size=HIDDEN_SIZE,
intermediate_size=INTERMEDIATE_SIZE, max_num_tokens=NUM_TOKENS,
top_k=TOP_K, device=DEVICE,
)
runner.l1_fp4 = l1_fp4; runner.l1_sf = l1_sf; runner.l1_gs = l1_gs_list
# Set L2 weights too (needed for _ensure_stacked)
l2_fp4, l2_sf, l2_gs_list = [], [], []
for e in expert_indices:
sf32 = fsf.float()
sf32[:, :INTERMEDIATE_SIZE] *= (ggs / mgs)
sf32[:, INTERMEDIATE_SIZE:] *= (ugs / mgs)
fsf = sf32.to(torch.float8_e4m3fn)
l1_fp4.append(fw); l1_sf.append(fsf); l1_gs.append(mgs)
dk = f"layers.{LAYER_IDX}.mlp.experts.{e}.down_proj.weight"
if dk in nvfp4_tensors:
dw = nvfp4_tensors[dk].to(DEVICE)
dsf = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.down_proj.weight_scale"].to(DEVICE)
dgs = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.down_proj.weight_scale_2"].item()
l2_fp4.append(dw.view(torch.float4_e2m1fn_x2).permute(1,0).contiguous())
l2_sf.append(dsf.permute(1,0).contiguous()); l2_gs_list.append(dgs)
l2_sf.append(dsf.permute(1,0).contiguous()); l2_gs.append(dgs)
else:
l2_fp4.append(torch.zeros(INTERMEDIATE_SIZE//2, HIDDEN_SIZE, dtype=torch.float4_e2m1fn_x2, device=DEVICE))
l2_sf.append(torch.ones(INTERMEDIATE_SIZE//16, HIDDEN_SIZE, dtype=torch.float8_e4m3fn, device=DEVICE))
l2_gs_list.append(1.0)
runner.l2_fp4 = l2_fp4; runner.l2_sf = l2_sf; runner.l2_gs = l2_gs_list
runner._ensure_stacked()
# Just use the runner's scale assembly
l1_gsa = torch.full((NUM_EXPERTS,), l1_gs, dtype=torch.float32, device=DEVICE)
l1_scale_a = runner._assemble_scales_cudagraph_safe(
slot_x_sf, expert_offsets[:NUM_EXPERTS+1],
padded_expert_offsets,
runner._padded_x_sf_buf_l1, runner._per_expert_scale_bufs_l1
l2_gs.append(1.0)
runner = CuTeDSLMoERunner(
num_experts=NUM_EXPERTS, hidden_size=HIDDEN_SIZE,
intermediate_size=INTERMEDIATE_SIZE, max_num_tokens=NUM_TOKENS,
top_k=TOP_K, device=DEVICE,
)
runner.l1_fp4 = l1_fp4; runner.l1_sf = l1_sf; runner.l1_gs = l1_gs
runner.l2_fp4 = l2_fp4; runner.l2_sf = l2_sf; runner.l2_gs = l2_gs
runner.set_swiglu_limit(SWIGLU_LIMIT)
l1_out = run_nvfp4_grouped_gemm(
mat_a=padded_x_fp4, mat_b=l1_mat_b,
scale_a=l1_scale_a, scale_b=l1_scale_b,
expert_offsets=padded_expert_offsets[1:NUM_EXPERTS+1],
global_scale_a=l1_gsa, global_scale_b=l1_gsb,
)
print(f" L1 out: shape={l1_out.shape} amax={l1_out.amax().item():.4f}")
print(f" L1 out NaN: {torch.isnan(l1_out).any().item()}")
with torch.no_grad():
runner.compute_activation_global_scales(hidden_states, topk_weights, topk_ids)
runner_out = runner.run(hidden_states, topk_weights, topk_ids)
# Extract real tokens
l1_out_real = l1_out[padded_dst]
print(f" L1 real: amax={l1_out_real.amax().item():.4f}")
print(f"Runner: amax={runner_out.amax().item():.4f}")
print(f"NaN: {torch.isnan(runner_out).any().item()}")
# BF16 reference L1
ref_l1 = torch.zeros(num_slots, 2*INTERMEDIATE_SIZE, dtype=torch.bfloat16, device=DEVICE)
for i, e in enumerate(expert_indices):
start = expert_offsets[i].item()
end = expert_offsets[i+1].item()
if start == end:
continue
x = slot_hidden[start:end]
gw = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.gate_proj.weight"].to(DEVICE)
uw = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.up_proj.weight"].to(DEVICE)
gsf = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.gate_proj.weight_scale"].to(DEVICE)
usf = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.up_proj.weight_scale"].to(DEVICE)
ggs = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.gate_proj.weight_scale_2"].item()
ugs = nvfp4_tensors[f"layers.{LAYER_IDX}.mlp.experts.{e}.up_proj.weight_scale_2"].item()
gate = x @ dequantize_nvfp4_weight(gw, gsf, ggs).T
up = x @ dequantize_nvfp4_weight(uw, usf, ugs).T
ref_l1[start:end] = torch.cat([gate, up], dim=1)
cos = F.cosine_similarity(ref_out.flatten().unsqueeze(0), runner_out.flatten().unsqueeze(0)).item()
mse = (ref_out - runner_out).pow(2).mean().item()
print(f"\nCosine: {cos:.6f} MSE: {mse:.6e}")
# Compare L1
cos_l1 = F.cosine_similarity(ref_l1.flatten().unsqueeze(0), l1_out_real.flatten().unsqueeze(0)).item()
print(f"\n L1 cosine vs BF16: {cos_l1:.6f}")
for t in range(NUM_TOKENS):
ct = F.cosine_similarity(ref_out[t].unsqueeze(0), runner_out[t].unsqueeze(0)).item()
print(f" Token {t}: cosine={ct:.4f}")
# Per-expert L1 comparison
for i in list(range(NUM_EXPERTS))[:5]:
start = expert_offsets[i].item()
end = expert_offsets[i+1].item()
if start == end:
continue
c = F.cosine_similarity(ref_l1[start:end].flatten().unsqueeze(0),
l1_out_real[start:end].flatten().unsqueeze(0)).item()
print(f" Expert {i} L1: cosine={c:.6f} ref_amax={ref_l1[start:end].amax().item():.4f} run_amax={l1_out_real[start:end].amax().item():.4f}")
if cos >= 0.98:
print(f"\n✅ PASS")
elif cos >= 0.90:
print(f"\n⚠️ MARGINAL")
else:
print(f"\n❌ FAIL")
if __name__ == "__main__":