252 lines
10 KiB
Python
252 lines
10 KiB
Python
#!/usr/bin/env python3
|
|
"""Verify single-layer attention output matches a PyTorch reference.
|
|
|
|
This script:
|
|
1. Loads layer 0 weights
|
|
2. Processes one token through the attention sub-block
|
|
3. Compares with a simple PyTorch reference
|
|
4. Identifies where the outputs diverge
|
|
"""
|
|
import os, sys, json, math, torch
|
|
from pathlib import Path
|
|
|
|
CHECKPOINT_DIR = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
|
|
LAYER_IDX = 0
|
|
|
|
# Correct E2M1 magnitudes
|
|
FP4_LUT = torch.tensor([0., 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0])
|
|
|
|
def dequant_nvfp4_weight(weight, weight_scale, weight_scale_2):
|
|
"""Dequantize NVFP4 weight to BF16."""
|
|
out_dim = weight.shape[0]
|
|
in_packed = weight.shape[1]
|
|
in_features = in_packed * 2
|
|
low = (weight & 0x0F).to(torch.int8)
|
|
high = (weight >> 4).to(torch.int8)
|
|
low_sign, low_idx = (low >> 3).bool(), (low & 0x07).long()
|
|
high_sign, high_idx = (high >> 3).bool(), (high & 0x07).long()
|
|
lut = FP4_LUT.to(device=weight.device, dtype=torch.float32)
|
|
low_f = lut[low_idx] * torch.where(low_sign, -1.0, 1.0)
|
|
high_f = lut[high_idx] * torch.where(high_sign, -1.0, 1.0)
|
|
w_f = torch.stack([low_f, high_f], dim=-1).reshape(out_dim, in_features)
|
|
scale_f = weight_scale.float() * weight_scale_2.float()
|
|
scale_expanded = scale_f.repeat_interleave(16, dim=1)
|
|
return (w_f * scale_expanded).bfloat16()
|
|
|
|
|
|
def nvfp4_linear(x, weight, weight_scale, weight_scale_2):
|
|
"""BF16 linear with NVFP4 dequant."""
|
|
w = dequant_nvfp4_weight(weight, weight_scale, weight_scale_2)
|
|
return torch.nn.functional.linear(x, w)
|
|
|
|
|
|
def rmsnorm(x, weight, eps=1e-6):
|
|
x_f = x.float()
|
|
rms = x_f.pow(2).mean(-1, keepdim=True).add(eps).rsqrt()
|
|
return (x_f * rms * weight.float()).bfloat16()
|
|
|
|
|
|
def main():
|
|
with open(os.path.join(CHECKPOINT_DIR, "config.json")) as f:
|
|
cfg = json.load(f)
|
|
|
|
n_h = cfg["num_attention_heads"] # 128
|
|
hd = cfg["head_dim"] # 512
|
|
H = cfg["hidden_size"] # 7168
|
|
rd = cfg.get("qk_rope_head_dim", cfg.get("rope_dim", 64)) # 64
|
|
o_rank = cfg.get("output_group_dim", 1024)
|
|
o_groups = cfg.get("num_output_groups", 16)
|
|
heads_per_group = n_h // o_groups # 8
|
|
group_input_dim = heads_per_group * hd # 4096
|
|
|
|
pre = f"model.layers.{LAYER_IDX}.self_attn"
|
|
|
|
# Load weights for this layer
|
|
from safetensors.torch import load_file
|
|
cdir = Path(CHECKPOINT_DIR)
|
|
with open(cdir / "model.safetensors.index.json") as f:
|
|
wm = json.load(f)["weight_map"]
|
|
|
|
w = {}
|
|
for key, shard in wm.items():
|
|
if key.startswith(f"model.layers.{LAYER_IDX}.self_attn.") and "compressor" not in key and "indexer" not in key:
|
|
data = load_file(str(cdir / shard))
|
|
w[key] = data[key].cuda()
|
|
|
|
print("Loaded attention weights:")
|
|
for k in sorted(w.keys()):
|
|
if "self_attn" in k:
|
|
print(f" ...{k.split('self_attn.')[1]}: {w[k].shape} {w[k].dtype}")
|
|
|
|
# Create input: random hidden state
|
|
torch.manual_seed(42)
|
|
x_raw = torch.randn(1, H, dtype=torch.bfloat16, device='cuda:0')
|
|
print(f"\nInput |x_raw| = {x_raw.abs().max():.4f}")
|
|
|
|
# RMSNorm
|
|
input_layernorm_w = None
|
|
for k in w:
|
|
if 'input_layernorm' in k:
|
|
input_layernorm_w = w[k]
|
|
break
|
|
if input_layernorm_w is not None:
|
|
x_normed = rmsnorm(x_raw, input_layernorm_w)
|
|
else:
|
|
x_normed = x_raw
|
|
print(f"After RMSNorm |x_normed| = {x_normed.abs().max():.4f}")
|
|
|
|
# === Q projection: q_a → q_a_norm → q_b ===
|
|
c_Q = nvfp4_linear(x_normed,
|
|
w[f"{pre}.q_a_proj.weight"],
|
|
w[f"{pre}.q_a_proj.weight_scale"],
|
|
w[f"{pre}.q_a_proj.weight_scale_2"])
|
|
print(f"\nc_Q: shape={c_Q.shape}, |c_Q|={c_Q.abs().max():.4f}")
|
|
|
|
# q_a_norm
|
|
if f"{pre}.q_a_norm.weight" in w:
|
|
c_Q = rmsnorm(c_Q, w[f"{pre}.q_a_norm.weight"])
|
|
print(f"After q_a_norm: |c_Q|={c_Q.abs().max():.4f}")
|
|
|
|
q = nvfp4_linear(c_Q,
|
|
w[f"{pre}.q_b_proj.weight"],
|
|
w[f"{pre}.q_b_proj.weight_scale"],
|
|
w[f"{pre}.q_b_proj.weight_scale_2"])
|
|
print(f"q: shape={q.shape}, |q|={q.abs().max():.4f}")
|
|
|
|
q_heads = q.reshape(1, n_h, hd)
|
|
print(f"q_heads: shape={q_heads.shape}")
|
|
|
|
# === KV projection + kv_norm ===
|
|
kv = nvfp4_linear(x_normed,
|
|
w[f"{pre}.kv_proj.weight"],
|
|
w[f"{pre}.kv_proj.weight_scale"],
|
|
w[f"{pre}.kv_proj.weight_scale_2"])
|
|
print(f"\nkv: shape={kv.shape}, |kv|={kv.abs().max():.4f}")
|
|
|
|
if f"{pre}.kv_norm.weight" in w:
|
|
kv = rmsnorm(kv, w[f"{pre}.kv_norm.weight"])
|
|
print(f"After kv_norm: |kv|={kv.abs().max():.4f}")
|
|
|
|
kv_heads = kv.reshape(1, 1, hd) # 1 KV head
|
|
print(f"kv_heads: shape={kv_heads.shape}")
|
|
|
|
# === Apply RoPE ===
|
|
half = rd // 2
|
|
freqs = 1.0 / (10000.0 ** (torch.arange(0, rd, 2, dtype=torch.float32, device='cuda') / rd))
|
|
pos = torch.tensor([0], dtype=torch.long, device='cuda')
|
|
cos = torch.cos(pos.float().unsqueeze(1) * freqs.unsqueeze(0)).bfloat16() # (1, half)
|
|
sin = torch.sin(pos.float().unsqueeze(1) * freqs.unsqueeze(0)).bfloat16()
|
|
|
|
def apply_rope(x, cos, sin):
|
|
nope = hd - rd
|
|
out = x.clone()
|
|
x_rope = x[:, :, nope:]
|
|
out[:, :, nope:][..., 0::2] = x_rope[..., 0::2] * cos - x_rope[..., 1::2] * sin
|
|
out[:, :, nope:][..., 1::2] = x_rope[..., 0::2] * sin + x_rope[..., 1::2] * cos
|
|
return out
|
|
|
|
def apply_inverse_rope(o, cos, sin):
|
|
nope = hd - rd
|
|
out = o.clone()
|
|
o_rope = o[:, :, nope:]
|
|
out[:, :, nope:][..., 0::2] = o_rope[..., 0::2] * cos + o_rope[..., 1::2] * sin
|
|
out[:, :, nope:][..., 1::2] = -o_rope[..., 0::2] * sin + o_rope[..., 1::2] * cos
|
|
return out
|
|
|
|
q_roped = apply_rope(q_heads, cos.unsqueeze(0), sin.unsqueeze(0))
|
|
kv_roped = apply_rope(kv_heads, cos.unsqueeze(0), sin.unsqueeze(0))
|
|
|
|
# === Attention (single KV entry → output = V) ===
|
|
# For 1 KV entry, attention output = V (softmax of scalar = 1)
|
|
# With K=V (both RoPE'd), output = V_roped
|
|
# Then inverse RoPE should recover kv (pre-RoPE)
|
|
attn_out = kv_roped # (1, 1, hd) — just V
|
|
|
|
attn_out_inv = apply_inverse_rope(attn_out, cos.unsqueeze(0), sin.unsqueeze(0))
|
|
|
|
# Check inverse RoPE recovery
|
|
diff = (attn_out_inv[0, 0].float() - kv_heads[0, 0].float()).abs().max()
|
|
print(f"\nInverse RoPE recovery: max diff = {diff:.6f} (should be ~0)")
|
|
|
|
# === Output projection ===
|
|
# For GQA, the attention output is (n_h, T, hd)
|
|
# Each Q head attended to the same KV, producing its own output
|
|
# For this test with 1 KV entry, all heads produce the same V
|
|
# In practice, each head has different Q, so different attention weights
|
|
# Let's use the actual multi-head attention output
|
|
|
|
# Proper multi-head attention with SDPA
|
|
q_input = q_roped # (1, n_h, hd) = (1, 128, 512)
|
|
k_input = kv_roped.expand(n_h, -1, -1) # (n_h, 1, hd) = (128, 1, 512)
|
|
v_input = kv_roped.expand(n_h, -1, -1)
|
|
|
|
attn_out_full = torch.nn.functional.scaled_dot_product_attention(
|
|
q_input, k_input, v_input,
|
|
scale=1.0 / math.sqrt(hd), is_causal=False) # (1, n_h, hd)
|
|
# Wait, shapes are (1, 128, 512) for q but (128, 1, 512) for k/v
|
|
# Need to fix: q is (1, n_h, hd) → permute to (n_h, 1, hd)
|
|
q_input = q_roped.permute(1, 0, 2) # (n_h, 1, hd) = (128, 1, 512)
|
|
k_input = kv_roped.squeeze(0).expand(n_h, -1, -1) # (n_h, 1, hd)
|
|
v_input = kv_roped.squeeze(0).expand(n_h, -1, -1) # (n_h, 1, hd)
|
|
|
|
attn_out_full = torch.nn.functional.scaled_dot_product_attention(
|
|
q_input, k_input, v_input,
|
|
scale=1.0 / math.sqrt(hd), is_causal=False) # (n_h, 1, hd)
|
|
attn_out_full = attn_out_full.permute(1, 0, 2) # (1, n_h, hd)
|
|
|
|
# Inverse RoPE
|
|
attn_out_inv = apply_inverse_rope(attn_out_full, cos.unsqueeze(0), sin.unsqueeze(0))
|
|
print(f"\nMulti-head attn output: shape={attn_out_inv.shape}, |attn|={attn_out_inv.abs().max():.4f}")
|
|
|
|
# Compare: per-head output should be close to kv (since 1 KV entry)
|
|
for h in [0, 1, 63, 127]:
|
|
diff = (attn_out_inv[0, h].float() - kv_heads[0, 0].float()).abs().max()
|
|
print(f" Head {h}: max diff from kv = {diff:.6f}")
|
|
|
|
attn_flat = attn_out_inv.reshape(1, n_h * hd) # (1, 65536)
|
|
print(f"\nattn_flat: shape={attn_flat.shape}, |attn_flat|={attn_flat.abs().max():.4f}")
|
|
|
|
# wo_a: grouped linear
|
|
attn_grouped = attn_flat.reshape(1, o_groups, heads_per_group * hd) # (1, 16, 4096)
|
|
oa_w = w[f"{pre}.o_a_proj.weight"].bfloat16() # (16384, 4096)
|
|
oa_3d = oa_w.reshape(o_groups, o_rank, group_input_dim) # (16, 1024, 4096)
|
|
attn_for_bmm = attn_grouped.permute(1, 0, 2) # (16, 1, 4096)
|
|
grouped_out = torch.bmm(attn_for_bmm, oa_3d.transpose(1, 2)) # (16, 1, 1024)
|
|
grouped_flat = grouped_out.permute(1, 0, 2).reshape(1, o_groups * o_rank) # (1, 16384)
|
|
|
|
print(f"grouped_flat: shape={grouped_flat.shape}, |grouped_flat|={grouped_flat.abs().max():.4f}")
|
|
|
|
# wo_b
|
|
F_attn = nvfp4_linear(grouped_flat,
|
|
w[f"{pre}.o_b_proj.weight"],
|
|
w[f"{pre}.o_b_proj.weight_scale"],
|
|
w[f"{pre}.o_b_proj.weight_scale_2"])
|
|
print(f"F_attn: shape={F_attn.shape}, |F_attn|={F_attn.abs().max():.4f}")
|
|
|
|
# === Sanity checks ===
|
|
print(f"\n{'='*50}")
|
|
print(f"ATTENTION SUB-BLOCK SUMMARY (Layer {LAYER_IDX})")
|
|
print(f"{'='*50}")
|
|
print(f"Input |x_normed| = {x_normed.abs().max():.4f}")
|
|
print(f"Q latent |c_Q| = {c_Q.abs().max():.4f}")
|
|
print(f"Q heads |q| = {q.abs().max():.4f}")
|
|
print(f"KV |kv| = {kv.abs().max():.4f}")
|
|
print(f"Attn out |attn_inv| = {attn_out_inv.abs().max():.4f}")
|
|
print(f"Grouped |grouped| = {grouped_flat.abs().max():.4f}")
|
|
print(f"F_attn (output) |F| = {F_attn.abs().max():.4f}")
|
|
print(f"Scale ratio F/x_norm = {F_attn.abs().max()/max(x_normed.abs().max(), 1e-8):.4f}")
|
|
|
|
# Check: is F_attn reasonable?
|
|
if F_attn.abs().max() > 100:
|
|
print(f"\n⚠️ WARNING: F_attn is very large ({F_attn.abs().max():.1f}). "
|
|
f"This will cause residual growth in the full model.")
|
|
elif F_attn.abs().max() < 0.01:
|
|
print(f"\n⚠️ WARNING: F_attn is very small. "
|
|
f"Attention output is being suppressed.")
|
|
else:
|
|
print(f"\n✅ F_attn is on a reasonable scale.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|