Files
nvfp4-megamoe-kernel/tests/test_residual_diagnostic.py

265 lines
12 KiB
Python

#!/usr/bin/env python3
"""Diagnostic: track per-layer magnitudes to find residual explosion.
Runs single token "The" through all 61 layers and prints:
|X_in|, |x_normed|, |F_attn|, |X_mid|, |F_ffn|, |X_next|
for each layer.
This identifies WHERE the residual stream starts exploding.
Usage (on B200):
source /root/dsv4-nvfp4-workspace/venv/bin/activate
cd /root/dsv4-nvfp4-workspace/kernel
python3 tests/test_residual_diagnostic.py
"""
import os, sys, math, json
import torch
from pathlib import Path
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
CHECKPOINT_DIR = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
NUM_GPUS = 8
# Import helpers from the main test
from tests.test_minimal_e2e import (
FP4_LUT, dequant_nvfp4_weight, nvfp4_linear, RMSNorm,
build_rope_cache, apply_rope_partial, apply_inverse_rope,
load_weights_to_cpu, get_layer_weights
)
from single_shot_inference import mHCBlock
def main():
device = 'cuda:0'
with open(os.path.join(CHECKPOINT_DIR, "config.json")) as f:
cfg = json.load(f)
n_layers = cfg["num_hidden_layers"]
H = cfg["hidden_size"]
n_h = cfg["num_attention_heads"]
hd = cfg["head_dim"]
rd = cfg.get("qk_rope_head_dim", cfg.get("rope_dim", 64))
n_hc = 4
o_rank = cfg.get("output_group_dim", 1024)
o_groups = cfg.get("num_output_groups", 16)
heads_per_group = n_h // o_groups
group_input_dim = heads_per_group * hd
print("Loading weights to CPU...")
all_weights = load_weights_to_cpu(CHECKPOINT_DIR)
embed_w = all_weights.get("model.embed_tokens.weight")
embed = torch.nn.Embedding.from_pretrained(embed_w.bfloat16().to(device))
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_DIR)
input_ids = tokenizer.encode("The")
tid = torch.tensor([input_ids[0]], dtype=torch.long, device=device)
positions = torch.tensor([0], dtype=torch.long, device=device)
emb = embed(tid)
from dsv4.layers.mhc import mHCLayer
X = mHCLayer.init_state(emb, n_hc)
print(f"\n{'L':>3} {'|X_in|':>10} {'|x_norm|':>10} {'|F_attn|':>10} {'|X_mid|':>10} {'|F_ffn|':>10} {'|X_out|':>10} nan? inf?")
for li in range(n_layers):
gpu = li % NUM_GPUS
dev = f"cuda:{gpu}"
X = X.to(dev)
torch.cuda.set_device(gpu)
w = get_layer_weights(all_weights, li, dev)
rope_cos, rope_sin = build_rope_cache(8192, rd, dev)
positions_dev = positions.to(dev)
# Build per-layer components
attn_mhc = mHCBlock(hidden_dim=H, n_hc=n_hc, device=dev)
attn_mhc.load_from_checkpoint(
all_weights[f"model.layers.{li}.attn_hc.fn"],
all_weights[f"model.layers.{li}.attn_hc.base"],
all_weights[f"model.layers.{li}.attn_hc.scale"])
ffn_mhc = mHCBlock(hidden_dim=H, n_hc=n_hc, device=dev)
ffn_mhc.load_from_checkpoint(
all_weights[f"model.layers.{li}.ffn_hc.fn"],
all_weights[f"model.layers.{li}.ffn_hc.base"],
all_weights[f"model.layers.{li}.ffn_hc.scale"])
attn_norm = RMSNorm(H, eps=cfg.get('rms_norm_eps', 1e-6), device=dev)
attn_norm.weight = all_weights[f"model.layers.{li}.input_layernorm.weight"].to(device=dev, dtype=torch.float32)
ffn_norm = RMSNorm(H, eps=cfg.get('rms_norm_eps', 1e-6), device=dev)
ffn_norm.weight = all_weights[f"model.layers.{li}.post_attention_layernorm.weight"].to(device=dev, dtype=torch.float32)
# Track magnitudes
x_in_mag = X.float().abs().max().item()
# ATTENTION
x_in_attn, attn_ctx = attn_mhc.pre_block(X)
x_normed = attn_norm.forward(x_in_attn)
x_normed_mag = x_normed.float().abs().max().item()
pre = f"model.layers.{li}.self_attn"
# Q projection
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"])
q_norm_w = w.get(f"{pre}.q_a_norm.weight")
if q_norm_w is not None:
c_Q_f = c_Q.float()
c_Q = (c_Q_f * c_Q_f.pow(2).mean(-1, keepdim=True).add(1e-6).rsqrt() * q_norm_w.float()).bfloat16()
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"])
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"])
kv_norm_w = w.get(f"{pre}.kv_norm.weight")
if kv_norm_w is not None:
kv_f = kv.float()
kv = (kv_f * kv_f.pow(2).mean(-1, keepdim=True).add(1e-6).rsqrt() * kv_norm_w.float()).bfloat16()
q_heads = q.reshape(1, n_h, hd)
kv_new = kv.reshape(1, 1, hd)
q_roped = apply_rope_partial(q_heads, positions_dev, rope_cos, rope_sin, hd, rd)
kv_roped = apply_rope_partial(kv_new, positions_dev, rope_cos, rope_sin, hd, rd)
k_exp = kv_roped.expand(n_h, -1, -1).contiguous()
v_exp = kv_roped.expand(n_h, -1, -1).contiguous()
q_input = q_roped.permute(1, 0, 2)
scale = 1.0 / math.sqrt(hd)
sink_key = f"{pre}.sinks"
if sink_key in w:
sinks = w[sink_key].to(device=dev)
sink_k = torch.zeros(n_h, 1, hd, dtype=torch.bfloat16, device=dev)
sink_v = torch.zeros(n_h, 1, hd, dtype=torch.bfloat16, device=dev)
k_ws = torch.cat([k_exp, sink_k], dim=1)
v_ws = torch.cat([v_exp, sink_v], dim=1)
sink_bias = torch.zeros(n_h, 1, 2, dtype=torch.bfloat16, device=dev)
for h in range(n_h):
sink_bias[h, :, -1] = sinks[h]
attn_out = torch.nn.functional.scaled_dot_product_attention(
q_input, k_ws, v_ws, attn_mask=sink_bias, scale=scale)
else:
attn_out = torch.nn.functional.scaled_dot_product_attention(
q_input, k_exp, v_exp, scale=scale, is_causal=False)
attn_out = attn_out.permute(1, 0, 2)
attn_out = apply_inverse_rope(attn_out, positions_dev, rope_cos, rope_sin, hd, rd)
# Output projection
attn_flat = attn_out.reshape(1, n_h * hd)
attn_grouped = attn_flat.reshape(1, o_groups, heads_per_group * hd)
oa_w = w[f"{pre}.o_a_proj.weight"].bfloat16()
oa_3d = oa_w.reshape(o_groups, o_rank, group_input_dim)
grouped_out = torch.bmm(attn_grouped.permute(1, 0, 2), oa_3d.transpose(1, 2))
grouped_flat = grouped_out.permute(1, 0, 2).reshape(1, o_groups * o_rank)
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"])
f_attn_mag = F_attn.float().abs().max().item()
X_mid = attn_mhc.post_block(X, F_attn, attn_ctx)
x_mid_mag = X_mid.float().abs().max().item()
# FFN (shared expert + routed, using the FIXED MoE loop)
x_ffn, ffn_ctx = ffn_mhc.pre_block(X_mid)
x_ffn_normed = ffn_norm.forward(x_ffn)
# Routed MoE
n_experts = cfg["n_routed_experts"]
top_k = cfg.get("num_experts_per_tok", 6)
routed_scaling = cfg.get("routed_scaling_factor", 2.5)
swiglu_limit = cfg.get("swiglu_limit", 10.0)
is_hash = li < 3
if is_hash:
tid2eid_key = f"model.layers.{li}.mlp.gate.tid2eid"
if tid2eid_key in w:
tid2eid = w[tid2eid_key]
tid_val = tid.item() if tid.device == dev else tid.to(dev).item()
expert_ids = tid2eid[tid_val]
expert_weights = torch.ones(top_k, dtype=torch.float32, device=dev) / top_k
else:
is_hash = False
if not is_hash:
gate_w = w[f"model.layers.{li}.mlp.gate.weight"]
logits = torch.nn.functional.linear(x_ffn_normed, gate_w.bfloat16())
activated = torch.sqrt(torch.nn.functional.softplus(logits.float()) + 1e-6)
e_bias_key = f"model.layers.{li}.mlp.gate.e_bias"
if e_bias_key in w:
activated = activated + w[e_bias_key].float().unsqueeze(0)
scores, indices = activated.topk(top_k, dim=-1)
unbiased = torch.sqrt(torch.nn.functional.softplus(logits.float()) + 1e-6)
unbiased_scores = torch.gather(unbiased, -1, indices)
expert_weights = unbiased_scores / unbiased_scores.sum(dim=-1, keepdim=True)
expert_ids = indices[0]
expert_outputs = []
for i, eid in enumerate(expert_ids):
eid_int = eid.item()
epre = f"model.layers.{li}.mlp.experts.{eid_int}"
gate = nvfp4_linear(x_ffn_normed, w[f"{epre}.gate_proj.weight"],
w[f"{epre}.gate_proj.weight_scale"], w[f"{epre}.gate_proj.weight_scale_2"])
up = nvfp4_linear(x_ffn_normed, w[f"{epre}.up_proj.weight"],
w[f"{epre}.up_proj.weight_scale"], w[f"{epre}.up_proj.weight_scale_2"])
silu_out = torch.nn.functional.silu(gate.float())
if swiglu_limit is not None:
silu_out = silu_out.clamp(-swiglu_limit, swiglu_limit)
up_clamped = up.float().clamp(-swiglu_limit, swiglu_limit)
else:
up_clamped = up.float()
hidden = (silu_out * up_clamped).bfloat16()
down = nvfp4_linear(hidden, w[f"{epre}.down_proj.weight"],
w[f"{epre}.down_proj.weight_scale"], w[f"{epre}.down_proj.weight_scale_2"])
expert_outputs.append(down)
routed_out = torch.zeros_like(x_ffn_normed)
for i, (out, wt) in enumerate(zip(expert_outputs, expert_weights)):
w_val = wt.item() if wt.dim() == 0 else wt[i].item() if wt.dim() == 1 else wt.flatten()[i].item()
routed_out = routed_out + (out.float() * w_val).bfloat16()
routed_out = (routed_out.float() * routed_scaling).bfloat16()
# Shared expert
se_pre = f"model.layers.{li}.mlp.shared_experts"
gate = nvfp4_linear(x_ffn_normed, w[f"{se_pre}.gate_proj.weight"],
w[f"{se_pre}.gate_proj.weight_scale"], w[f"{se_pre}.gate_proj.weight_scale_2"])
up = nvfp4_linear(x_ffn_normed, w[f"{se_pre}.up_proj.weight"],
w[f"{se_pre}.up_proj.weight_scale"], w[f"{se_pre}.up_proj.weight_scale_2"])
hidden = (torch.nn.functional.silu(gate.float()).clamp(-10, 10) * up.float().clamp(-10, 10)).bfloat16()
shared_out = nvfp4_linear(hidden, w[f"{se_pre}.down_proj.weight"],
w[f"{se_pre}.down_proj.weight_scale"], w[f"{se_pre}.down_proj.weight_scale_2"])
F_ffn = routed_out + shared_out
f_ffn_mag = F_ffn.float().abs().max().item()
X_next = ffn_mhc.post_block(X_mid, F_ffn, ffn_ctx)
x_out_mag = X_next.float().abs().max().item()
has_nan = torch.isnan(X_next).any().item()
has_inf = torch.isinf(X_next).any().item()
print(f"{li:3d} {x_in_mag:10.3f} {x_normed_mag:10.3f} {f_attn_mag:10.3f} {x_mid_mag:10.3f} {f_ffn_mag:10.3f} {x_out_mag:10.3f} {'NaN' if has_nan else ''} {'INF' if has_inf else ''}")
X = X_next
del w
torch.cuda.empty_cache()
# Final logits
X = X.to('cuda:0')
lm_w = all_weights.get("lm_head.weight", embed_w).bfloat16().to('cuda:0')
final_norm_w = all_weights.get("model.norm.weight")
x_out = X[:, 0, :]
if final_norm_w is not None:
xf = x_out.float()
rms = xf.pow(2).mean(-1, keepdim=True).add(1e-6).rsqrt()
x_out = (xf * rms * final_norm_w.to('cuda:0').float()).bfloat16()
logits = torch.nn.functional.linear(x_out, lm_w)
top5_vals, top5_ids = torch.topk(logits[0], 5)
top5_str = ' '.join([f'{tokenizer.decode([t.item()])}({v.item():.1f})'
for t, v in zip(top5_ids, top5_vals)])
print(f"\nTop-5: {top5_str}")
if __name__ == "__main__":
main()