biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add degeneration test 2: falsify mHC residual growth root cause

Browse Source
This commit is contained in:
biondizzle
2026-06-03 08:18:01 +00:00
parent c77b83fffc
commit 11c2d5fe53
1 changed files with 450 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

450
tests/unit/test_degeneration_2_mhc_falsify.py Normal file
View File

@@ -0,0 +1,450 @@
#!/usr/bin/env python3
"""DEGENERATION TEST 2 — Falsify the mHC "root cause".
Claim under test: "|X|=860 compresses the logit range so the model can't distinguish tokens."
Why it's suspect: there is a final RMSNorm before the LM head, and RMSNorm is
scale-invariant — it divides the magnitude out. So |X|=860 and |X|=8 should produce
the SAME logits (modulo the learned norm weight). Also, the residual grows just as
much during prefill yet prefill/first-token is correct — magnitude common to both
phases cannot be what breaks only decode.
Procedure:
1. Confirm the final norm exists and is applied.
2. Falsification: compute logits with X as-is (|X|≈860) and X/100, compare.
If argmax matches and cos≈1.0 → mHC growth is EXONERATED.
If they differ → something downstream is magnitude-sensitive → norm is missing/broken.
This test loads the FULL model (61 layers, 8 GPUs, production values).
It runs one decode step and captures the final-layer residual for the comparison.
"""
import os, sys, time, json, math
import torch
import torch.nn.functional as F
# This test imports and reuses single_shot_inference.py's infrastructure
# but intercepts at the hc_head / final_norm / lm_head stage.
CHECKPOINT_DIR = os.environ.get("CHECKPOINT_DIR", "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4")
NUM_GPUS = 8
def main():
print("=" * 70)
print("DEGENERATION TEST 2 — Falsify mHC residual growth root cause")
print("=" * 70)
# We need to run the full pipeline to get the final-layer residual X.
# The simplest approach: import single_shot's main, but intercept at the lm_head.
# Instead of re-implementing everything, we'll modify the decode loop to capture X
# and do the comparison after one decode step.
# Load the model using single_shot's infrastructure
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from single_shot_inference import (
load_all_weights, build_rope_cache, rmsnorm, unweighted_rmsnorm,
mHCLayer, HcHead, KVCache, Compressor, Indexer,
make_nvfp4_linear, get_nvfp4_weight, do_nvfp4_linear_ref,
forward_layer, moe_forward, _cache_layer_weights_no_experts,
_load_moe_weights_stacked, _load_shared_expert_weights,
FP4_LUT, HC_EPS, THINK_START, THINK_END, USER_TOKEN, ASSISTANT_TOKEN,
kill_stale_gpu_processes,
)
from transformers import AutoTokenizer
from dsv4.layers.mhc import mHCLayer as mHCLayerProd
from dsv4.layers.router import Router
from dsv4.layers.moe import Nvfp4MoE
from dsv4.layers.shared_expert import Nvfp4SharedExpert
from dsv4.layers.linear import Nvfp4Linear
from dsv4.layers.grouped_linear import Nvfp4GroupedLinear
from dsv4.ops.quantize import quantize_weight_to_nvfp4
t0 = time.time(); torch.manual_seed(42)
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"]
hd = cfg["head_dim"]; n_h = cfg["num_attention_heads"]
rd = cfg.get("qk_rope_head_dim", 64)
cr = cfg.get("compress_ratios", [128] * n_layers)
PROMPT = "The capital of France is"
print(f"Model: {n_layers} layers, {n_h} heads, hd={hd}, rope_dim={rd}")
# Load weights
print(f"\nLoading weights..."); all_w = load_all_weights(CHECKPOINT_DIR)
# Build production components (same as single_shot main)
kill_stale_gpu_processes()
for g in range(NUM_GPUS): torch.cuda.set_device(g); torch.cuda.empty_cache()
torch.cuda.set_device(0)
# mHC + norms
attn_mhcs, ffn_mhcs, attn_norms, ffn_norms = {}, {}, {}, {}
for li in range(n_layers):
dev = f"cuda:{li % NUM_GPUS}"
for tag, blocks, fn_s, base_s, scale_s in [
("attn", attn_mhcs, f"model.layers.{li}.attn_hc.fn", f"model.layers.{li}.attn_hc.base", f"model.layers.{li}.attn_hc.scale"),
("ffn", ffn_mhcs, f"model.layers.{li}.ffn_hc.fn", f"model.layers.{li}.ffn_hc.base", f"model.layers.{li}.ffn_hc.scale"),
]:
fn, base, scale = all_w.get(fn_s), all_w.get(base_s), all_w.get(scale_s)
if fn is not None and base is not None and scale is not None:
m = mHCLayerProd(hidden_dim=H, n_hc=4, t_max_sinkhorn=20, device=dev)
n = 4
m.load_weights(
W_pre=fn[0:n].to(dev, torch.float32), W_post=fn[n:2*n].to(dev, torch.float32),
W_comb=fn[2*n:].to(dev, torch.float32),
S_pre=base[0:n].reshape(1, n).to(dev, torch.float32),
S_post=base[n:2*n].reshape(n, 1).to(dev, torch.float32),
S_comb=base[2*n:].reshape(n, n).to(dev, torch.float32),
alpha_pre=scale[0].item(), alpha_post=scale[1].item(), alpha_comb=scale[2].item(),
)
blocks[li] = m
an_k = f"model.layers.{li}.input_layernorm.weight"
if an_k in all_w: attn_norms[li] = all_w[an_k].to(dev, torch.float32)
fn_k = f"model.layers.{li}.post_attention_layernorm.weight"
if fn_k in all_w: ffn_norms[li] = all_w[fn_k].to(dev, torch.float32)
# Attention linears
prod_lins = {}
for li in range(n_layers):
dev = f"cuda:{li % NUM_GPUS}"; pfx = f"model.layers.{li}.self_attn"
torch.cuda.set_device(li % NUM_GPUS)
pl = {}
pl['q_a'] = make_nvfp4_linear(7168, 1536, dev, all_w, pfx, 'q_a_proj')
pl['q_b'] = make_nvfp4_linear(1536, 65536, dev, all_w, pfx, 'q_b_proj')
pl['kv'] = make_nvfp4_linear(7168, 512, dev, all_w, pfx, 'kv_proj')
n_local_groups = cfg.get('o_groups', 16)
heads_per_group = n_h // n_local_groups
o_rank_val = cfg.get('o_lora_rank', 1024)
wo_a = Nvfp4GroupedLinear(n_local_groups=n_local_groups, heads_per_group=heads_per_group,
head_dim=hd, o_lora_rank=o_rank_val, max_num_tokens=8192, device=dev)
oa_w_nvfp4, oa_ws, oa_ws2, oa_isc = get_nvfp4_weight(all_w, pfx, 'o_a_proj')
if oa_w_nvfp4 is not None and oa_ws is not None:
wo_a.load_nvfp4_weight(oa_w_nvfp4.to(dev), oa_ws.to(dev),
oa_ws2.to(dev) if oa_ws2 is not None else None,
oa_isc.to(dev) if oa_isc is not None else None)
else:
oa_bf = all_w.get(f"{pfx}.o_a_proj.weight")
if oa_bf is not None: wo_a.set_bf16_weight(oa_bf.bfloat16().to(dev))
pl['o_a'] = wo_a
wo_a._use_runtime_gsa = True
pl['o_b'] = make_nvfp4_linear(16384, 7168, dev, all_w, pfx, 'o_b_proj')
prod_lins[li] = pl
# Routers, MoE, shared experts
routers, moe_runners, se_runners = {}, {}, {}
for li in range(n_layers):
dev = f"cuda:{li % NUM_GPUS}"; pfx = f"model.layers.{li}.mlp"
torch.cuda.set_device(li % NUM_GPUS); torch.cuda.synchronize()
is_hash = (li < cfg.get("num_hash_layers", 3)) and (f"{pfx}.gate.tid2eid" in all_w)
router = Router(hidden_size=H, num_experts=cfg["n_routed_experts"],
top_k=cfg.get("num_experts_per_tok", 6),
routed_scaling_factor=cfg.get("routed_scaling_factor", 2.5),
mode="hash" if is_hash else "dense",
vocab_size=cfg.get("vocab_size", 128000) if is_hash else None, device=dev)
if is_hash:
router.load_weights(hash_lut=all_w[f"{pfx}.gate.tid2eid"].to(dev, torch.int32))
else:
eb = all_w.get(f"{pfx}.gate.e_score_correction_bias")
gate_w, gate_ws, gate_ws2, gate_isc = get_nvfp4_weight(all_w, pfx, 'gate')
E = cfg["n_routed_experts"]
if gate_w is not None and gate_ws is not None:
gate_lin = Nvfp4Linear(in_features=H, out_features=E, device=dev)
gate_w_view = gate_w.to(dev).view(torch.float4_e2m1fn_x2) if gate_w.dtype == torch.uint8 else gate_w.to(dev)
gate_lin.fp4 = [gate_w_view]
gate_lin.sf = [gate_ws.to(dev)]
ws2_v = gate_ws2.float().item() if gate_ws2 is not None else 1.0
isc_v = gate_isc.float().item() if gate_isc is not None else 1.0/(6.0*448.0)
gate_lin.gs = [1.0]
gate_lin.ws2 = [torch.tensor([ws2_v], device=dev, dtype=torch.float32)]
gate_lin._activation_global_scale = isc_v
gate_lin._use_runtime_gsa = True
gate_lin.finalize_weights()
router.load_nvfp4_gate(gate_lin)
router.load_weights(e_bias=eb.to(dev, torch.float32))
else:
gw = all_w.get(f"{pfx}.gate.weight")
if gw is not None:
g_bf16 = gw if gw.shape == (E, H) else gw.T.contiguous()
g_bf16 = g_bf16.bfloat16().to(dev)
from dsv4.ops.quantize import quantize_to_nvfp4
g_fp4, g_sf, g_gs = quantize_to_nvfp4(g_bf16)
gate_lin = Nvfp4Linear(in_features=H, out_features=E, device=dev)
gate_lin.fp4 = [g_fp4]; gate_lin.sf = [g_sf]; gate_lin.gs = [g_gs]
gate_lin.ws2 = [torch.tensor([g_gs], device=dev, dtype=torch.float32)]
gate_lin._activation_global_scale = 1.0 / (6.0 * 448.0)
gate_lin._use_runtime_gsa = True
gate_lin.finalize_weights()
router.load_nvfp4_gate(gate_lin)
router.load_weights(e_bias=eb.to(dev, torch.float32))
router.finalize_weights(); routers[li] = router
moe = Nvfp4MoE(num_experts=cfg["n_routed_experts"], hidden_size=H,
intermediate_size=cfg.get("moe_intermediate_size", 3072),
top_k=cfg.get("num_experts_per_tok", 6), device=dev)
moe.set_swiglu_limit(cfg.get("swiglu_limit", 10.0))
moe.set_fused_swiglu(True)
_load_moe_weights_stacked(all_w, li, pfx, dev, moe, cfg)
moe._ensure_stacked()
moe._use_runtime_gsa = True
moe_runners[li] = moe
se = Nvfp4SharedExpert(hidden_size=H, intermediate_size=cfg.get("moe_intermediate_size", 3072),
device=dev, swiglu_limit=cfg.get("swiglu_limit", 10.0))
se.set_fused_swiglu(True)
_load_shared_expert_weights(all_w, li, pfx, dev, se, cfg)
se._ensure_initialized()
se._use_runtime_gsa = True
se_runners[li] = se
if (li+1) % 10 == 0: print(f" Built {li+1}/{n_layers} MoE layers")
torch.cuda.empty_cache()
# Global weights
torch.cuda.set_device(0)
embed_w = all_w.get("model.embed_tokens.weight")
embed = torch.nn.Embedding.from_pretrained(embed_w.bfloat16().to('cuda:0'))
lm_w_raw = all_w.get("lm_head.weight", embed_w).bfloat16().to('cuda:0')
lm_head_lin = Nvfp4Linear(lm_w_raw.shape[1], lm_w_raw.shape[0], max_num_tokens=8192, device='cuda:0')
lm_fp4, lm_sf, lm_gs = quantize_weight_to_nvfp4(lm_w_raw.T.contiguous())
lm_head_lin.fp4 = [lm_fp4.permute(1, 0).contiguous()]
lm_head_lin.sf = [lm_sf.permute(1, 0).contiguous()]
lm_head_lin.gs = [lm_gs]
lm_head_lin.ws2 = [None]
lm_head_lin._activation_global_scale = 1.0 / (6.0 * 448.0)
lm_head_lin._use_runtime_gsa = True
lm_head_lin.finalize_weights()
final_norm_w = all_w.get("model.norm.weight")
if final_norm_w is not None:
final_norm_w = final_norm_w.to('cuda:0', torch.float32)
hc_head = HcHead(H, 4, 'cuda:0')
hc_fn = all_w.get("model.hc_head.hc_fn")
hc_base = all_w.get("model.hc_head.hc_base")
hc_scale = all_w.get("model.hc_head.hc_scale")
if hc_fn is not None and hc_base is not None:
hc_head.load(hc_fn, hc_base, hc_scale)
# RoPE
rp = cfg.get("rope_scaling", cfg.get("rope_parameters", {}))
rt = rp.get("type", rp.get("rope_type", "yarn")); rf = rp.get("factor", 16.0)
rtheta = cfg.get("rope_theta", 10000.)
romax = rp.get("original_max_position_embeddings", 65536)
rbfast, rbslow = rp.get("beta_fast", 32), rp.get("beta_slow", 1)
rope_caches = {g: build_rope_cache(romax, rd, f"cuda:{g}", rtheta, rt, rf, romax, rbfast, rbslow)
for g in range(NUM_GPUS)}
# KV caches, compressors, indexers
kv_caches, compressors, indexers = {}, {}, {}
n_ih = cfg.get("index_n_heads", 64); ihd = cfg.get("index_head_dim", 128)
itk = cfg.get("index_topk", 1024)
for li in range(n_layers):
dev = f"cuda:{li % NUM_GPUS}"; ratio = cr[li] if li < len(cr) else 128
max_comp = (8192 + ratio - 1) // ratio if ratio > 0 else 0
kv_caches[li] = KVCache(hd, cfg.get("sliding_window", 128), max_comp=max_comp, device=dev,
indexer_key_dim=ihd, compress_ratio=ratio, indexer_top_k=itk, rope_dim=rd)
if ratio > 0: compressors[li] = Compressor(ratio, hd, H, dev)
if ratio == 4: indexers[li] = Indexer(n_ih, ihd, itk, dev)
# Cache layer weights
devs = [f"cuda:{g}" for g in range(NUM_GPUS)]
layer_w = _cache_layer_weights_no_experts(all_w, n_layers, devs)
del all_w; import gc; gc.collect()
for g in range(NUM_GPUS): torch.cuda.set_device(g); torch.cuda.empty_cache()
torch.cuda.set_device(0)
for li in range(n_layers):
pfx = f"model.layers.{li}.self_attn.compressor"
if li in compressors: compressors[li].load(layer_w[li], pfx, dev=f"cuda:{li % NUM_GPUS}")
if li in indexers: indexers[li].load(layer_w[li], f"{pfx}.indexer", dev=f"cuda:{li % NUM_GPUS}")
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_DIR)
bos = tokenizer.bos_token_id or 0
input_ids = [bos, USER_TOKEN]
input_ids += tokenizer.encode('\n\n' + PROMPT, add_special_tokens=False)
input_ids.append(ASSISTANT_TOKEN)
input_ids.append(THINK_START)
print(f"\nPhase: Prefill + 1 decode step")
print(f" Input: {len(input_ids)} tokens")
# Prefill
PREFILL_CHUNK = 128
n_prefill = len(input_ids)
prefill_ids = torch.tensor(input_ids, dtype=torch.long, device='cuda:0')
prefill_ids32 = prefill_ids.to(torch.int32)
all_positions = torch.arange(n_prefill, dtype=torch.long, device='cuda:0')
chunk_starts = list(range(0, n_prefill, PREFILL_CHUNK))
X = None
for ci, cs in enumerate(chunk_starts):
ce = min(cs + PREFILL_CHUNK, n_prefill)
chunk_len = ce - cs
chunk_ids = prefill_ids[cs:ce]
chunk_ids32 = prefill_ids32[cs:ce]
chunk_positions = all_positions[cs:ce]
chunk_embed = embed(chunk_ids)
X = mHCLayerProd.init_state(chunk_embed)
for li in range(n_layers):
gpu = li % NUM_GPUS
if X.device != torch.device(f"cuda:{gpu}"): X = X.to(f"cuda:{gpu}")
torch.cuda.set_device(gpu)
X = forward_layer(X, layer_w[li], li, cfg, *rope_caches[gpu],
attn_mhcs.get(li), ffn_mhcs.get(li),
attn_norms.get(li), ffn_norms.get(li),
kv_caches[li], chunk_positions, chunk_ids32,
compressors.get(li), indexers.get(li),
moe_runners.get(li), se_runners.get(li), routers.get(li),
prod_lin=prod_lins.get(li))
X = X.to('cuda:0'); torch.cuda.set_device(0)
print(f" Chunk {ci+1}/{len(chunk_starts)}: OK", flush=True)
# Decode step 1
dec_tid = torch.tensor([input_ids[-1]], dtype=torch.long, device='cuda:0')
dec_tid32 = dec_tid.to(torch.int32)
dec_pos = torch.tensor([n_prefill - 1], dtype=torch.long, device='cuda:0')
X = mHCLayerProd.init_state(embed(dec_tid))
for li in range(n_layers):
gpu = li % NUM_GPUS
if X.device != torch.device(f"cuda:{gpu}"): X = X.to(f"cuda:{gpu}")
torch.cuda.set_device(gpu)
X = forward_layer(X, layer_w[li], li, cfg, *rope_caches[gpu],
attn_mhcs.get(li), ffn_mhcs.get(li),
attn_norms.get(li), ffn_norms.get(li),
kv_caches[li], dec_pos, dec_tid32,
compressors.get(li), indexers.get(li),
moe_runners.get(li), se_runners.get(li), routers.get(li),
prod_lin=prod_lins.get(li))
X = X.to('cuda:0'); torch.cuda.set_device(0)
torch.cuda.synchronize()
# ================================================================
# TEST 2: Falsification
# ================================================================
print(f"\n{'='*70}")
print("TEST 2 — Falsify mHC residual growth root cause")
print(f"{'='*70}")
# Step 1: Confirm final norm exists and is applied
print(f"\n1. FINAL NORM CHECK:")
print(f" final_norm_w exists: {final_norm_w is not None}")
if final_norm_w is not None:
print(f" final_norm_w shape: {final_norm_w.shape}, dtype: {final_norm_w.dtype}")
print(f" final_norm_w range: [{final_norm_w.min().item():.6f}, {final_norm_w.max().item():.6f}]")
else:
print(f" *** CRITICAL: final_norm_w is MISSING! This is likely the real bug! ***")
# Step 2: Trace the full path: X → hc_head → final_norm → lm_head → logits
print(f"\n2. RESIDUAL INSPECTION:")
X_abs_max = X.abs().max().item()
print(f" |X| (final layer residual) = {X_abs_max:.4f}")
print(f" X shape: {X.shape}, dtype: {X.dtype}")
# hc_head: takes (T, n_hc, d) → (T, d)
x_out = hc_head.forward(X) if hc_head is not None else X[:, 0, :]
x_out_max = x_out.abs().max().item()
print(f" |x_out| (after hc_head) = {x_out_max:.4f}")
print(f" x_out shape: {x_out.shape}, dtype: {x_out.dtype}")
# Apply final norm
if final_norm_w is not None:
x_normed = rmsnorm(x_out, final_norm_w)
x_normed_max = x_normed.abs().max().item()
print(f" |x_normed| (after final_norm) = {x_normed_max:.4f}")
# Verify scale invariance: rmsnorm should divide out magnitude
x_out_tiny = x_out / 100.0
x_normed_tiny = rmsnorm(x_out_tiny, final_norm_w)
cos_norm = F.cosine_similarity(x_normed.flatten().float(), x_normed_tiny.flatten().float(), dim=0).item()
print(f" RMSNorm scale invariance: cos(x_normed, x_normed_tiny) = {cos_norm:.8f}")
print(f" (Expected: 1.0 — RMSNorm is scale-invariant)")
else:
x_normed = x_out
print(f" *** NO FINAL NORM APPLIED — logits will be magnitude-dependent! ***")
# Step 3: Falsification — compute logits with X and X/100
print(f"\n3. FALSIFICATION: logits with |X|={X_abs_max:.1f} vs |X/100|={X_abs_max/100:.1f}")
# Path A: logits with X as-is
x_out_A = hc_head.forward(X) if hc_head is not None else X[:, 0, :]
if final_norm_w is not None:
x_out_A = rmsnorm(x_out_A, final_norm_w)
logits_A = lm_head_lin(x_out_A)
# Path B: logits with X scaled down by 100
X_scaled = X / 100.0
x_out_B = hc_head.forward(X_scaled) if hc_head is not None else X_scaled[:, 0, :]
if final_norm_w is not None:
x_out_B = rmsnorm(x_out_B, final_norm_w)
logits_B = lm_head_lin(x_out_B)
torch.cuda.synchronize()
logits_A_f = logits_A.float()
logits_B_f = logits_B.float()
argmax_A = logits_A_f.argmax().item()
argmax_B = logits_B_f.argmax().item()
cos_AB = F.cosine_similarity(logits_A_f.flatten(), logits_B_f.flatten(), dim=0).item()
top5_A_vals, top5_A_ids = logits_A_f.topk(5)
top5_B_vals, top5_B_ids = logits_B_f.topk(5)
print(f"\n logits_A (|X|={X_abs_max:.1f}):")
print(f" range: [{logits_A_f.min().item():.2f}, {logits_A_f.max().item():.2f}]")
print(f" argmax: {argmax_A} ('{tokenizer.decode([argmax_A])}')")
print(f" top-5: {[(tokenizer.decode([tid.item()]), f'{val.item():.2f}') for tid, val in zip(top5_A_ids, top5_A_vals)]}")
print(f"\n logits_B (|X/100|={X_abs_max/100:.2f}):")
print(f" range: [{logits_B_f.min().item():.2f}, {logits_B_f.max().item():.2f}]")
print(f" argmax: {argmax_B} ('{tokenizer.decode([argmax_B])}')")
print(f" top-5: {[(tokenizer.decode([tid.item()]), f'{val.item():.2f}') for tid, val in zip(top5_B_ids, top5_B_vals)]}")
print(f"\n cos(logits_A, logits_B) = {cos_AB:.8f}")
print(f" argmax_A == argmax_B: {argmax_A == argmax_B}")
# Step 4: Also check hc_head behavior — is it magnitude-sensitive?
print(f"\n4. HC_HEAD MAGNITUDE SENSITIVITY:")
# hc_head does: rmsnorm(X) → linear → sigmoid → sum * X
# The sigmoid step makes it potentially magnitude-sensitive
# Let's check: does hc_head(X) scale linearly with |X|?
x_out_A_raw = hc_head.forward(X) if hc_head is not None else X[:, 0, :]
x_out_B_raw = hc_head.forward(X / 100.0) if hc_head is not None else (X / 100.0)[:, 0, :]
cos_hc = F.cosine_similarity(x_out_A_raw.flatten().float(), (x_out_B_raw * 100.0).flatten().float(), dim=0).item()
print(f" cos(hc_head(X), hc_head(X/100)*100) = {cos_hc:.8f}")
print(f" (If hc_head is NOT magnitude-sensitive, this should be 1.0)")
print(f" |hc_head(X)| = {x_out_A_raw.abs().max().item():.4f}")
print(f" |hc_head(X/100)| = {x_out_B_raw.abs().max().item():.6f}")
print(f" |hc_head(X/100)*100| = {(x_out_B_raw * 100.0).abs().max().item():.4f}")
# Step 5: Final verdict
print(f"\n{'='*70}")
print("VERDICT:")
print(f"{'='*70}")
if final_norm_w is None:
print(" *** CRITICAL: FINAL NORM IS MISSING! ***")
print(" The model has no RMSNorm before the LM head.")
print(" This means logits are magnitude-dependent → mHC residual growth IS the problem.")
print(" FIX: Apply the final norm before lm_head.")
elif cos_AB >= 0.999:
print(" mHC residual growth is EXONERATED.")
print(f" cos(logits_A, logits_B) = {cos_AB:.8f} ≈ 1.0")
print(f" argmax_A={argmax_A}, argmax_B={argmax_B}")
print(" |X| magnitude does NOT affect logits (RMSNorm divides it out).")
print(" The degeneration cause is elsewhere — likely Test 1 (chat template).")
elif argmax_A != argmax_B:
print(" mHC residual growth IS magnitude-sensitive despite final norm.")
print(f" argmax_A={argmax_A} ≠ argmax_B={argmax_B}")
print(f" cos = {cos_AB:.8f}")
print(" Something downstream of the residual is magnitude-sensitive.")
print(" Check: hc_head linearity, lm_head quantization, or the final norm is misapplied.")
else:
print(f" Inconclusive: argmax matches but cos={cos_AB:.8f} < 0.999")
print(" Logits are similar but not identical at different |X| scales.")
print(" The magnitude does have SOME effect, but may not be the primary cause.")
print(f"{'='*70}")
if __name__ == "__main__":
main()