Files
colibri/c/tools/make_glm_oracle.py
T
Vincenzo b04c2ca50e Merge pull request #322 from woolcoxm/fix/oracle-transformers-pin
oracle: hard-fail on transformers < 5.11.0 (interleaved-RoPE floor, #281)
2026-07-16 19:46:20 +02:00

169 lines
7.7 KiB
Python

"""Costruisce un GLM-5.2 (glm_moe_dsa) MINUSCOLO a pesi random come ORACOLO.
Architettura vera (MLA + DSA indexer + router sigmoid/noaux_tc + shared expert),
dimensioni minuscole. Salva pesi+config in c/glm_tiny/ e un riferimento greedy in
c/ref_glm.json. seq corta (<= index_topk) cosi' il DSA seleziona tutte le key e
l'attenzione coincide con la MLA densa: il motore C puo' validare senza implementare
l'indexer sparso.
--fp8: salva i pesi come FP8 e4m3 + scale a blocchi 128x128 (layout del checkpoint reale
GLM-5.2-FP8) invece di bf16, cosi' convert_fp8_to_int4.py puo' esercitare il path FP8->int4
su un modello minuscolo. PRIMA di calcolare ref_glm.json fa il round-trip dei pesi per FP8
(quant->dequant, copy_ nel modello): cosi' il riferimento riflette ESATTAMENTE il modello
FP8 che il converter legge, non il modello bf16 a precisione piena. Default: bf16 (oracolo
originale invariato).
EN: --fp8 writes FP8 e4m3 + 128x128 block scale_inv (real GLM-5.2-FP8 layout) instead of bf16,
EN: so convert_fp8_to_int4.py can run its FP8->int4 path on a tiny model. ref_glm.json is
EN: computed AFTER the FP8 round-trip, so the reference matches exactly what the converter
EN: ingests. Default: bf16 (original oracle unchanged)."""
import json, sys, argparse
from pathlib import Path
# --- Version gate (must run BEFORE the heavy `from transformers import ...`,
# which triggers transformers' lazy-loading and can in turn reset the in-memory
# __version__ attribute; importlib.metadata reads the installed package version
# directly and is immune to that). ---
#
# GLM-5.2's MLA attention uses interleaved (DeepSeek-style) RoPE, which is what
# the C engine implements. transformers < 5.11.0 applied split-half (Llama-style)
# RoPE in GlmMoeDsa* instead; an oracle built on those versions drifts and the
# engine then scores 25/32 instead of the documented 32/32 (issue #281). Weights
# come out identical across versions — only the forward pass differs — so there
# is no safe "partial" run: a too-old transformers silently produces an invalid
# ref_glm.json. Hard-fail rather than warn. EN: same.
import transformers
from importlib.metadata import version as _pkg_version, PackageNotFoundError
_MIN_TRANSFORMERS = (5, 11)
def _tf_version_tuple():
try:
v = _pkg_version("transformers") # authoritative: installed dist metadata
except PackageNotFoundError:
v = getattr(transformers, "__version__", "0") # fallback (editable/src installs)
out = []
for part in v.split(".")[:2]: # major.minor only
out.append(int("".join(c for c in part if c.isdigit()) or "0"))
while len(out) < 2:
out.append(0)
return tuple(out[:2])
_tf_ver = _tf_version_tuple()
if _tf_ver < _MIN_TRANSFORMERS:
sys.exit(
f"\nERROR: make_glm_oracle.py requires transformers >= "
f"{'.'.join(map(str, _MIN_TRANSFORMERS))}.0 (found {_tf_ver[0]}.{_tf_ver[1]}). "
f"GLM-5.2 MLA uses interleaved RoPE; older versions apply split-half RoPE "
f"and silently produce an oracle the engine scores 25/32 against (issue #281). "
f"Upgrade: pip install -U 'transformers>=5.11'\n"
)
import torch
from transformers import GlmMoeDsaConfig, GlmMoeDsaForCausalLM
sys.path.insert(0, str(Path(__file__).resolve().parent)) # importa glm_fp8_emit se lanciato da c/
from glm_fp8_emit import (fp8_block_quantize, fp8_block_dequantize, keep_f32,
save_fp8_safetensors, unfuse_experts)
ap = argparse.ArgumentParser()
ap.add_argument("--fp8", action="store_true",
help="salva in FP8 e4m3 + 128x128 block scale_inv (layout GLM-5.2-FP8) e "
"calcola ref_glm.json sul modello dopo il round-trip FP8. "
"EN: write FP8 e4m3 + block scale_inv, ref computed on FP8-rounded model")
args = ap.parse_args()
torch.manual_seed(1234)
cfg = GlmMoeDsaConfig(
vocab_size=256,
hidden_size=128,
intermediate_size=64, # MLP densa (primi 3 layer)
moe_intermediate_size=32, # expert
num_hidden_layers=5, # 3 densi + 2 sparse
first_k_dense_replace=3,
num_attention_heads=4,
num_key_value_heads=4,
n_routed_experts=8,
num_experts_per_tok=2,
n_shared_experts=1,
q_lora_rank=64,
kv_lora_rank=32,
qk_nope_head_dim=24,
qk_rope_head_dim=8, # pari -> interleave ok; head_dim diventa 8
v_head_dim=32,
index_topk=4096, # >> seq_len -> DSA seleziona tutto (no-op)
index_head_dim=16,
index_n_heads=2,
n_group=1,
topk_group=1,
norm_topk_prob=True,
routed_scaling_factor=2.5,
rope_parameters={"rope_type": "default", "rope_theta": 10000.0},
tie_word_embeddings=False,
rms_norm_eps=1e-5,
attention_bias=False,
max_position_embeddings=4096,
)
cfg._attn_implementation = "eager"
model = GlmMoeDsaForCausalLM(cfg).eval()
# rende i pesi non banali (default init e' molto piccolo): scala router/bias per topk vario
with torch.no_grad():
for n, p in model.named_parameters():
if p.dim() >= 2:
p.normal_(0, 0.05)
# bias di correzione del router: valori distinti cosi' la selezione e' sensata
for i, layer in enumerate(model.model.layers):
if hasattr(layer.mlp, "gate"):
layer.mlp.gate.e_score_correction_bias.copy_(
torch.linspace(-0.1, 0.1, cfg.n_routed_experts))
# --fp8: round-trip dei pesi quantizzabili per FP8 PRIMA di calcolare il riferimento,
# cosi' ref_glm.json riflette esattamente il modello FP8 che il converter leggera'.
# Norme/router/bias (keep_f32) restano a precisione piena. EN: --fp8: round-trip quantizable
# weights through FP8 before computing the reference, so ref_glm.json matches the FP8 model.
if args.fp8:
with torch.no_grad():
for n, p in model.named_parameters():
if keep_f32(n, p) or p.dim() != 2:
continue
q, s = fp8_block_quantize(p)
p.copy_(fp8_block_dequantize(q, s))
print("=== state_dict tensors (names used by the C loader) ===")
for n, p in model.state_dict().items():
print(f" {n:60s} {tuple(p.shape)}")
prompt = [3, 14, 159, 26, 53, 58, 200, 11, 77, 240, 5, 99] # token id arbitrari, seq corta
ids = torch.tensor([prompt])
with torch.no_grad():
out = model.generate(ids, max_new_tokens=20, do_sample=False, use_cache=True)
full = out[0].tolist()
print("\nprompt:", prompt)
print("full :", full)
# teacher-forcing: un singolo forward su tutta la sequenza -> argmax per posizione.
# Per il greedy vale tf_pred[i] == full[i+1] per i >= len(prompt)-1; serve a validare
# il PREFILL del motore C separandolo dal decode.
with torch.no_grad():
lg = model(torch.tensor([full]), use_cache=False).logits[0] # [seq, vocab]
tf_pred = lg.argmax(-1).tolist()
print("tf_pred:", tf_pred)
# Unfuse experts AFTER reference generation (model needs fused weights for
# forward/generate) but BEFORE saving — the real checkpoint and the converter
# + C engine all expect per-expert 2-D gate_proj/up_proj/down_proj tensors.
sd = model.state_dict()
unfuse_experts(sd)
Path("glm_tiny").mkdir(parents=True, exist_ok=True) # safetensors/json won't create the dir themselves
if args.fp8:
n_fp8, n_tot = save_fp8_safetensors(sd, "glm_tiny/model.safetensors")
print(f"\nsaved FP8: {n_fp8} e4m3 tensors (+{n_tot - n_fp8} scale_inv sidecars / f32) "
f"-> glm_tiny/model.safetensors")
else:
from safetensors.torch import save_file
save_file({k: v.contiguous() for k, v in sd.items()}, "glm_tiny/model.safetensors")
json.dump(cfg.to_dict(), open("glm_tiny/config.json", "w"))
json.dump({"prompt_ids": prompt, "full_ids": full, "tf_pred": tf_pred}, open("ref_glm.json", "w"))
print("saved: glm_tiny/ (weights + config) and ref_glm.json"
+ (" [fp8]" if args.fp8 else ""))