Rewrite test: diagnose whether warmup gs matters at inference time

This commit is contained in:
2026-05-19 07:49:41 +00:00
parent 496848e158
commit 04ad6409e5

View File

@@ -2,21 +2,19 @@
"""
Reproduce the vLLM empty-output bug outside the container.
Runs the FULL model forward pass: embedding → 61 decoder layers → LM head.
Uses CuTeDSL NVFP4 runners for quantized layers, BF16 matmuls for others.
Strategy: Run the model in FULL BF16 (dequantized weights) and compare
against CuTeDSL at each projection. Also check: does the warmup gs
cause issues at inference time?
Compares two approaches:
A) Warmup gs (what vLLM does) — 1 token random sample per layer
B) Dynamic gs (compute per-batch) — uses quantize_to_nvfp4 each call
If A produces garbage and B produces reasonable output, the warmup gs is wrong.
Key diagnostic: inspect CuTeDSL runner.run() to see if it uses the
fixed warmup gs or recomputes per-call.
Usage (on B200):
source /root/nvfp4-megamoe-kernel/tests/.venv/bin/activate
python3 tests/test_model_forward_b200.py
"""
import sys, os, json, torch, torch.nn.functional as F, time
import sys, os, json, torch, torch.nn.functional as F, inspect
from safetensors import safe_open
REPO = "/root/nvfp4-megamoe-kernel"
@@ -24,24 +22,9 @@ sys.path.insert(0, REPO)
MODEL = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
DEV = "cuda:0"
# Model config
H = 7168
NH = 128
HD = 512
NOPE = 448
ROPE = 64
QL = 1536
OL = 1024
OG = 16
HPG = NH // OG
HC = 4
SL = 10.0
H = 7168; NH = 128; HD = 512; NOPE = 448; ROPE = 64
QL = 1536; OL = 1024; OG = 16; HPG = NH // OG
EPS = 1e-6
INTER = 3072
N_EXPERTS = 384
TOP_K = 6
N_LAYERS = 61
VOCAB = 129280
E2M1 = torch.tensor([0,.5,1.,1.5,2.,3.,4.,6.,-0,-.5,-1.,-1.5,-2.,-3.,-4.,-6.], dtype=torch.float32)
@@ -83,246 +66,170 @@ def make_runner(w, sf, gs_t, inf, outf, fused=False, lw=None):
r.finalize_weights(); r._ensure_initialized()
return r
def cosim(a, b):
return F.cosine_similarity(a.flatten().unsqueeze(0).float(), b.flatten().unsqueeze(0).float().to(a.device)).item()
class Layer0Runner:
"""Runs layer 0 forward with CuTeDSL kernels."""
def __init__(self, wm, model_dir, use_warmup_gs=True):
G = lambda k: P(k, wm, model_dir).to(DEV)
p = "model.layers.0"; a = f"{p}.self_attn"; m = f"{p}.mlp"
# Attention
self.qa_w = G(f"{a}.q_a_proj.weight"); self.qa_sf = G(f"{a}.q_a_proj.weight_scale"); self.qa_gs = G(f"{a}.q_a_proj.weight_scale_2")
self.qb_w = G(f"{a}.q_b_proj.weight"); self.qb_sf = G(f"{a}.q_b_proj.weight_scale"); self.qb_gs = G(f"{a}.q_b_proj.weight_scale_2")
self.kv_w = G(f"{a}.kv_proj.weight"); self.kv_sf = G(f"{a}.kv_proj.weight_scale"); self.kv_gs = G(f"{a}.kv_proj.weight_scale_2")
self.woa = G(f"{a}.o_a_proj.weight")
self.wob_w = G(f"{a}.o_b_proj.weight"); self.wob_sf = G(f"{a}.o_b_proj.weight_scale"); self.wob_gs = G(f"{a}.o_b_proj.weight_scale_2")
self.qn = G(f"{a}.q_a_norm.weight"); self.kvn = G(f"{a}.kv_norm.weight")
self.anorm = G(f"{p}.input_layernorm.weight"); self.fnorm = G(f"{p}.post_attention_layernorm.weight")
# Compressor
self.ckv_w = G(f"{a}.compressor.kv_proj.weight"); self.ckv_sf = G(f"{a}.compressor.kv_proj.weight_scale"); self.ckv_gs = G(f"{a}.compressor.kv_proj.weight_scale_2")
self.cg_w = G(f"{a}.compressor.gate_proj.weight"); self.cg_sf = G(f"{a}.compressor.gate_proj.weight_scale"); self.cg_gs = G(f"{a}.compressor.gate_proj.weight_scale_2")
# MHC
self.hca_fn = G(f"{p}.attn_hc.fn"); self.hcf_fn = G(f"{p}.ffn_hc.fn")
self.hca_b = G(f"{p}.attn_hc.base"); self.hcf_b = G(f"{p}.ffn_hc.base")
self.hca_s = G(f"{p}.attn_hc.scale"); self.hcf_s = G(f"{p}.ffn_hc.scale")
# Create runners
self.r_qa = make_runner(self.qa_w, self.qa_sf, self.qa_gs, self.qa_w.shape[1]*2, self.qa_w.shape[0])
self.r_qb = make_runner(self.qb_w, self.qb_sf, self.qb_gs, self.qb_w.shape[1]*2, self.qb_w.shape[0])
self.r_kv = make_runner(self.kv_w, self.kv_sf, self.kv_gs, self.kv_w.shape[1]*2, self.kv_w.shape[0])
self.r_wob = make_runner(self.wob_w, self.wob_sf, self.wob_gs, self.wob_w.shape[1]*2, self.wob_w.shape[0])
self.r_ckv = make_runner(self.ckv_w, self.ckv_sf, self.ckv_gs, self.ckv_w.shape[1]*2, self.ckv_w.shape[0])
self.r_cg = make_runner(self.cg_w, self.cg_sf, self.cg_gs, self.cg_w.shape[1]*2, self.cg_w.shape[0])
self.use_warmup_gs = use_warmup_gs
if use_warmup_gs:
# Warmup with 1 token (what vLLM does)
with torch.no_grad():
d = torch.randn(1, H, dtype=torch.bfloat16, device=DEV)*2.0
self.r_qa.compute_activation_global_scale(d)
self.r_kv.compute_activation_global_scale(d)
self.r_ckv.compute_activation_global_scale(d)
self.r_cg.compute_activation_global_scale(d)
d2 = torch.randn(1, QL, dtype=torch.bfloat16, device=DEV)*2.0
self.r_qb.compute_activation_global_scale(d2)
d3 = torch.randn(1, OG*OL, dtype=torch.bfloat16, device=DEV)*2.0
self.r_wob.compute_activation_global_scale(d3)
def forward_projection(self, x, runner, in_features, name):
"""Run a single NVFP4 projection, optionally recomputing gs."""
if not self.use_warmup_gs:
from cutedsl.bridge import quantize_activation_nvfp4
# Dynamic gs: recompute for this specific input
amax = x.amax().item()
gs = amax / (6.0 * 448.0) if amax > 0 else 1.0 / 2688.0
runner._activation_global_scale = gs
return runner.run(x)
def main():
torch.cuda.set_device(0)
torch.manual_seed(42)
print("=" * 70)
print(" Full Model Forward Test: Reproduce vLLM Empty Output")
print(" Diagnose: Why does vLLM produce empty output?")
print("=" * 70)
with open(os.path.join(MODEL, "model.safetensors.index.json")) as f:
wm = json.load(f)["weight_map"]
G = lambda k: P(k, wm, MODEL).to(DEV)
# ── Load embedding ────────────────────────────────────────────────
print("\n--- Loading embedding layer ---")
emb_key = "model.embed_tokens.weight"
emb = G(emb_key)
print(f" embed_tokens: {emb.shape} dtype={emb.dtype}")
# ── INSPECT: How does CuTeDSL runner.run() use gs? ────────────────
print("\n--- INSPECTING CuTeDSL runner internals ---")
from cutedsl.nvfp4_linear import CuTeDSLNvfp4Linear
from cutedsl.bridge import quantize_activation_nvfp4
# ── Load LM head ──────────────────────────────────────────────────
lm_head_key = "lm_head.weight"
if lm_head_key in wm:
lm_head = G(lm_head_key)
else:
lm_head = emb # tied weights
print(f" lm_head: {lm_head.shape}")
print("\n quantize_activation_nvfp4 signature:")
sig = inspect.signature(quantize_activation_nvfp4)
print(f" {sig}")
# ── Load final norm ───────────────────────────────────────────────
fnorm_key = "model.norm.weight"
fnorm_w = G(fnorm_key)
print(f" final_norm: {fnorm_w.shape}")
print("\n CuTeDSLNvfp4Linear._run_impl source (key lines):")
src = inspect.getsource(CuTeDSLNvfp4Linear._run_impl)
for i, line in enumerate(src.split('\n')):
stripped = line.strip()
if any(kw in stripped for kw in ['global_scale', '_activation', 'quantize', 'return', 'def ']):
print(f" L{i}: {stripped}")
# ── Token IDs for "The capital of France is" ──────────────────────
# DeepSeek V3/V4 uses a Llama-style BPE tokenizer
# Use token IDs that we know work. If tokenizer isn't available,
# just use token 0,1,2,3,4 as a test — we're checking for
# garbage output (all NaN or all same logit), not text quality.
# ── CRITICAL TEST: warmup gs vs per-input gs ──────────────────────
print("\n--- CRITICAL TEST: warmup gs vs per-input gs ---")
p = "model.layers.0"; a = f"{p}.self_attn"
qa_w = G(f"{a}.q_a_proj.weight"); qa_sf = G(f"{a}.q_a_proj.weight_scale"); qa_gs = G(f"{a}.q_a_proj.weight_scale_2")
# Load embedding + norm weight
emb = G("model.embed_tokens.weight")
anorm = G(f"{p}.input_layernorm.weight")
# Create runner with warmup gs
r = make_runner(qa_w, qa_sf, qa_gs, qa_w.shape[1]*2, qa_w.shape[0])
torch.manual_seed(42)
warmup = torch.randn(1, H, dtype=torch.bfloat16, device=DEV)*2.0
with torch.no_grad():
r.compute_activation_global_scale(warmup)
print(f" Warmup gs (random input amax={warmup.amax():.4f}): {r._activation_global_scale:.8f}")
# Get REAL input (embedding output)
token_ids = torch.tensor([1, 450, 8403, 315, 5413, 374], dtype=torch.long, device=DEV)
print(f" token_ids: {token_ids.tolist()}")
NT = len(token_ids)
# ── Embed ─────────────────────────────────────────────────────────
print("\n--- Running embedding lookup ---")
with torch.no_grad():
hidden = emb[token_ids] # (NT, H)
print(f" hidden: {hidden.shape} amax={hidden.amax():.4f} NaN={torch.isnan(hidden).any()}")
hidden = emb[token_ids]
normed = rms(hidden, anorm, EPS)
print(f" Real input (after RMS norm) amax: {normed.amax():.4f}")
# ── Create layer 0 runner (warmup gs, like vLLM) ─────────────────
print("\n--- Creating layer 0 runner (warmup gs) ---")
layer0 = Layer0Runner(wm, MODEL, use_warmup_gs=True)
# What gs would the real input need?
real_gs = normed.amax().item() / (6.0 * 448.0)
print(f" Correct gs for real input: {real_gs:.8f}")
print(f" Ratio warmup/correct: {r._activation_global_scale / real_gs:.4f}" if real_gs > 0 else " real_gs is 0!")
# ── Run layer 0 attention projections ─────────────────────────────
print("\n--- Running layer 0 attention (CuTeDSL, warmup gs) ---")
# Run with warmup gs
with torch.no_grad():
normed = rms(hidden, layer0.anorm, EPS)
print(f" normed: amax={normed.amax():.4f} NaN={torch.isnan(normed).any()}")
out_warmup = r.run(normed)
qa_out = layer0.r_qa.run(normed)
print(f" q_a: amax={qa_out.amax():.4f} NaN={torch.isnan(qa_out).any()}")
kv_out = layer0.r_kv.run(normed)
print(f" kv: amax={kv_out.amax():.4f} NaN={torch.isnan(kv_out).any()}")
# q_a norm → q_b
qa_normed = rms(qa_out, layer0.qn, EPS)
qb_out = layer0.r_qb.run(qa_normed)
print(f" q_b: amax={qb_out.amax():.4f} NaN={torch.isnan(qb_out).any()}")
# ── Compare with BF16 reference ───────────────────────────────────
print("\n--- Comparing layer 0 q_a with BF16 reference ---")
qa_bf16 = normed @ dequant(layer0.qa_w, layer0.qa_sf, layer0.qa_gs.item()).T
c = cosim(qa_out, qa_bf16)
print(f" q_a cosine (warmup gs): {c:.6f} {'' if c>=0.98 else ''}")
# ── Now test with DYNAMIC gs (recomputed per input) ───────────────
print("\n--- Testing with dynamic gs (per-input) ---")
# Create a fresh runner and compute gs from the actual input
r_qa2 = make_runner(layer0.qa_w, layer0.qa_sf, layer0.qa_gs, layer0.qa_w.shape[1]*2, layer0.qa_w.shape[0])
# Run with dynamic gs (recompute for this input)
r2 = make_runner(qa_w, qa_sf, qa_gs, qa_w.shape[1]*2, qa_w.shape[0])
with torch.no_grad():
r_qa2.compute_activation_global_scale(normed)
qa_out2 = r_qa2.run(normed)
c2 = cosim(qa_out2, qa_bf16)
print(f" q_a cosine (dynamic gs): {c2:.6f} {'' if c2>=0.98 else ''}")
r2.compute_activation_global_scale(normed)
out_dynamic = r2.run(normed)
# ── Test the FULL model: layer 0 only, then check LM head ─────────
print("\n--- Full forward: layer 0 → LM head ---")
# BF16 reference
qa_bf16 = dequant(qa_w, qa_sf, qa_gs.item())
with torch.no_grad():
ref = normed @ qa_bf16.T
c_warmup = F.cosine_similarity(out_warmup.flatten().unsqueeze(0).float(), ref.flatten().unsqueeze(0).float()).item()
c_dynamic = F.cosine_similarity(out_dynamic.flatten().unsqueeze(0).float(), ref.flatten().unsqueeze(0).float()).item()
print(f"\n q_a cosine vs BF16 (warmup gs): {c_warmup:.6f} {'' if c_warmup>=0.98 else ''}")
print(f" q_a cosine vs BF16 (dynamic gs): {c_dynamic:.6f} {'' if c_dynamic>=0.98 else ''}")
print(f" amax warmup: {out_warmup.amax():.4f} amax dynamic: {out_dynamic.amax():.4f} amax ref: {ref.amax():.4f}")
# ── Test: run FULL model in BF16 (1 layer) then check logits ──────
print("\n--- FULL BF16 model: 1 layer → LM head ---")
lm_head = G("lm_head.weight")
fnorm_w = G("model.norm.weight")
qn = G(f"{a}.q_a_norm.weight")
kvn = G(f"{a}.kv_norm.weight")
fnorm_l0 = G(f"{p}.post_attention_layernorm.weight")
# Dequantize all layer 0 attention weights
qa_bf16 = dequant(qa_w, qa_sf, qa_gs.item())
qb_w = G(f"{a}.q_b_proj.weight"); qb_sf = G(f"{a}.q_b_proj.weight_scale"); qb_gs = G(f"{a}.q_b_proj.weight_scale_2")
kv_bf16 = dequant(G(f"{a}.kv_proj.weight"), G(f"{a}.kv_proj.weight_scale"), G(f"{a}.kv_proj.weight_scale_2").item())
qb_bf16 = dequant(qb_w, qb_sf, qb_gs.item())
woa = G(f"{a}.o_a_proj.weight") # already BF16
wob_bf16 = dequant(G(f"{a}.o_b_proj.weight"), G(f"{a}.o_b_proj.weight_scale"), G(f"{a}.o_b_proj.weight_scale_2").item())
# Simple layer 0 forward (attention only, no MoE for speed)
with torch.no_grad():
x = hidden.clone()
normed = rms(x, layer0.anorm, EPS)
print(f" Input: amax={x.amax():.4f}")
# Attention projections
qa = layer0.r_qa.run(normed)
kv = layer0.r_kv.run(normed)
qa_n = rms(qa, layer0.qn, EPS)
qb = layer0.r_qb.run(qa_n)
# RMS norm
x = rms(x, anorm, EPS)
# Skip actual attention (FlashMLA not available) — just use
# a random attention output to test the wo_a → wo_b path
o = torch.randn(NT, NH, HD, dtype=torch.bfloat16, device=DEV) * 0.1
# Attention projections (BF16)
qa = x @ qa_bf16.T
kv = x @ kv_bf16.T
qa_n = rms(qa, qn, EPS)
qb = qa_n @ qb_bf16.T
print(f" q_a: amax={qa.amax():.4f}, kv: amax={kv.amax():.4f}, q_b: amax={qb.amax():.4f}")
# wo_a: BF16 BMM
woa = layer0.woa
o_2d = o.reshape(NT, NH * HD)
z = o_2d @ woa.T
z2 = z.reshape(NT, OG, -1)
# Skip attention, use random output
o = torch.randn(len(token_ids), NH, HD, dtype=torch.bfloat16, device=DEV) * 0.1
# Simpler: just check if the wo_b projection works
z_flat = torch.randn(NT, OG * OL, dtype=torch.bfloat16, device=DEV) * 2.0
wob_out = layer0.r_wob.run(z_flat)
print(f" wo_b output: amax={wob_out.amax():.4f} NaN={torch.isnan(wob_out).any()}")
# wo_a: BMM (o_a_proj is (OG*OL, HPG*HD))
o_grouped = o.view(len(token_ids), OG, HPG * HD).permute(1, 0, 2)
woa_3d = woa.view(OG, OL, HPG * HD)
z = torch.bmm(o_grouped, woa_3d.transpose(1, 2)).permute(1, 0, 2).reshape(len(token_ids), OG * OL)
# ── Now run LM head on the hidden state ───────────────────────────
print("\n--- LM head (BF16 matmul) ---")
with torch.no_grad():
normed_final = rms(x, fnorm_w, EPS)
logits = normed_final @ lm_head.T # (NT, VOCAB)
print(f" logits: {logits.shape} amax={logits.amax():.4f} NaN={torch.isnan(logits).any()}")
# wo_b
attn_out = z @ wob_bf16.T
print(f" attn_out (BF16): amax={attn_out.amax():.4f}")
# Skip MoE, just add residual
x = hidden + attn_out
# Final norm + LM head
x_normed = rms(x, fnorm_w, EPS)
logits = x_normed @ lm_head.T
print(f" logits: amax={logits.amax():.4f} NaN={torch.isnan(logits).any()}")
# Check if logits are reasonable
top5 = torch.topk(logits[-1], 5)
print(f" top5 token IDs: {top5.indices.tolist()}")
print(f" top5 IDs: {top5.indices.tolist()}")
print(f" top5 logits: {[f'{v:.2f}' for v in top5.values.tolist()]}")
# Check logit variance (garbage = all same or extreme values)
log_std = logits[-1].float().std().item()
log_range = (logits[-1].float().amax() - logits[-1].float().amin()).item()
print(f" logit std: {log_std:.4f} range: {log_range:.4f}")
if log_std < 0.01:
print(" ❌ LOGITS ARE FLAT — model is producing garbage!")
elif log_std > 100:
print(" ❌ LOGITS ARE EXPLODED — model is producing garbage!")
else:
print(" ✅ Logits look reasonable for a single layer test")
print(f" logit std: {log_std:.4f}")
# ── Key diagnostic: warmup gs vs actual gs ───────────────────────
# ── KEY INSIGHT: check if the runner re-reads gs at inference time
print("\n" + "=" * 70)
print(" DIAGNOSTIC: Warmup gs vs Actual gs")
print(" KEY: Does runner.run() use FIXED warmup gs or RECOMPUTE?")
print("=" * 70)
# What gs did warmup compute?
print(f" r_qa warmup gs: {layer0.r_qa._activation_global_scale:.8f}")
print(f" r_kv warmup gs: {layer0.r_kv._activation_global_scale:.8f}")
print(f" r_wob warmup gs: {layer0.r_wob._activation_global_scale:.8f}")
# What gs would the actual input produce?
# Monkey-patch the gs and see if output changes
r3 = make_runner(qa_w, qa_sf, qa_gs, qa_w.shape[1]*2, qa_w.shape[0])
with torch.no_grad():
actual_gs_qa = normed.amax().item() / (6.0 * 448.0)
actual_gs_kv = normed.amax().item() / (6.0 * 448.0)
print(f" actual gs for q_a input: {actual_gs_qa:.8f}")
print(f" ratio warmup/actual for q_a: {layer0.r_qa._activation_global_scale / actual_gs_qa:.4f}" if actual_gs_qa > 0 else " actual gs is 0!")
r3.compute_activation_global_scale(normed)
gs_original = r3._activation_global_scale
out_original = r3.run(normed).clone()
# The KEY question: does the runner use warmup gs at inference time,
# or does quantize_activation_nvfp4 recompute it?
print("\n--- How does CuTeDSL runner.run() use gs? ---")
from cutedsl.nvfp4_linear import CuTeDSLNvfp4Linear
import inspect
run_src = inspect.getsource(CuTeDSLNvfp4Linear.run)
# Check if it references _activation_global_scale
if '_activation_global_scale' in run_src:
print(" run() uses _activation_global_scale (FIXED from warmup)")
# Change gs by 10x
r3._activation_global_scale = gs_original * 10.0
out_changed = r3.run(normed).clone()
c_changed = F.cosine_similarity(out_original.flatten().unsqueeze(0).float(), out_changed.flatten().unsqueeze(0).float()).item()
print(f" Original gs: {gs_original:.8f}")
print(f" Changed gs: {gs_original * 10:.8f}")
print(f" Cosine sim after 10x gs change: {c_changed:.6f}")
if abs(c_changed - 1.0) < 0.001:
print(" ➡️ Changing gs has NO effect on output!")
print(" ➡️ The runner recomputes gs internally at inference time.")
print(" ➡️ Warmup gs is IRRELEVANT — the bug is elsewhere.")
else:
print(" run() does NOT use _activation_global_scale")
# Check quantize_activation_nvfp4
from cutedsl.bridge import quantize_activation_nvfp4
qsrc = inspect.getsource(quantize_activation_nvfp4)
if 'global_scale' in qsrc:
print(" quantize_activation_nvfp4 accepts global_scale as parameter")
if '_activation_global_scale' in qsrc:
print(" quantize_activation_nvfp4 reads _activation_global_scale")
# Check _run_impl
run_impl_src = inspect.getsource(CuTeDSLNvfp4Linear._run_impl)
print(f"\n _run_impl length: {len(run_impl_src)} chars")
# Find where gs is used
for i, line in enumerate(run_impl_src.split('\n')):
if 'global_scale' in line or '_activation' in line:
print(f" _run_impl line {i}: {line.strip()}")
print(" ➡️ Changing gs DOES change the output!")
print(" ➡️ The runner uses the warmup gs at inference time.")
print(" ➡️ Wrong warmup gs would cause wrong quantization → garbage.")
if __name__ == "__main__":