01_export_quant.py

#!/usr/bin/env python
"""
01_export_quant.py
Gemma 4 E4B (gfp78/gemma4-bund-merged)  ->  ONNX-Decoder  ->  q4f16

Phasen:
  A) Modell laden (bf16 -> fp32 auf CPU/GPU)
  B) Cache-Layer + Head-Dims per Trockenlauf ermitteln (NICHT raten)
  C) torch.onnx.export (Legacy-Tracer, dynamo=False)
  D) Konsolidierung zu EINER .onnx_data (finaler Name direkt!)
  E) MatMul4BitsQuantizer -> q4f16
  F) Verifikation + Groessen-Check

Laufzeit-Schaetzung: A-D ~40 Min, E ~10-25 Min.
"""

import gc
import os
import shutil
import sys
from pathlib import Path

import torch
import onnx

os.environ.setdefault("HF_HOME", "/root/hf")

MODEL_ID = "gfp78/gemma4-bund-merged"
OUT = Path("/root/train/gemma4-bund-onnx")
OUT.mkdir(parents=True, exist_ok=True)

FP32_ONNX = OUT / "decoder_model_merged.onnx"
FP32_DATA = "decoder_model_merged.onnx_data"      # relativer Name, PFLICHT
Q4_ONNX = OUT / "decoder_model_merged_q4f16.onnx"
Q4_DATA = "decoder_model_merged_q4f16.onnx_data"


def log(msg):
    print(f"\n=== {msg}", flush=True)


# ---------------------------------------------------------------- A) laden
log("A) Modell laden")
from transformers import AutoTokenizer, AutoModelForImageTextToText, DynamicCache

tok = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID, dtype=torch.float32, device_map="cpu"
)
model.eval()

# Sprach-Turm + lm_head separat: language_model liefert nur last_hidden_state,
# die lm_head-Projektion sitzt eine Ebene hoeher.
lm = model.model.language_model
lm_head = model.lm_head if hasattr(model, "lm_head") else model.get_output_embeddings()
tcfg = lm.config
HIDDEN = tcfg.hidden_size
print("hidden_size:", HIDDEN)


# ------------------------------------------- B) Cache-Geometrie ermitteln
log("B) Trockenlauf: Anzahl Cache-Layer + Head-Dims ermitteln")
with torch.no_grad():
    probe_ids = torch.tensor([[1, 2, 3, 4]], dtype=torch.long)
    probe_emb = lm.get_input_embeddings()(probe_ids)
    probe = lm(inputs_embeds=probe_emb, use_cache=True, return_dict=True)

pkv = probe.past_key_values
N_CACHE = len(pkv.layers)          # transformers 5.13: .layers, nicht .key_cache
print("n_cache_layers:", N_CACHE)  # erwartet: 24 (nicht 42!)

# Gemma 4 hat ZWEI Head-Dims: 256 (sliding window) und 512 (full attention).
# Deshalb pro Layer die echte Form auslesen statt eine globale anzunehmen.
KV_SHAPES = []
for i in range(N_CACHE):
    k = pkv.layers[i].keys
    KV_SHAPES.append((int(k.shape[1]), int(k.shape[3])))  # (n_kv_heads, head_dim)
for i, s in enumerate(KV_SHAPES):
    print(f"  layer {i:2d}: n_kv_heads={s[0]}, head_dim={s[1]}")

del probe, pkv, probe_emb
gc.collect()


# ------------------------------------------------------------ C) Wrapper
class DecoderWrapper(torch.nn.Module):
    """inputs_embeds + attention_mask + position_ids + past -> logits + present"""

    def __init__(self, lm, lm_head, n_cache):
        super().__init__()
        self.lm = lm
        self.lm_head = lm_head
        self.n_cache = n_cache

    def forward(self, inputs_embeds, attention_mask, position_ids, *past):
        cache = None
        if len(past) == 2 * self.n_cache and past[0].shape[2] > 0:
            cache = DynamicCache(config=self.lm.config)
            for i in range(self.n_cache):
                cache.update(past[2 * i], past[2 * i + 1], i)

        out = self.lm(
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=cache,
            use_cache=True,
            return_dict=True,
        )
        logits = self.lm_head(out.last_hidden_state)

        present = []
        for i in range(self.n_cache):
            present.append(out.past_key_values.layers[i].keys)
            present.append(out.past_key_values.layers[i].values)
        return (logits, *present)


wrapper = DecoderWrapper(lm, lm_head, N_CACHE).eval()

log("C) Dummy-Inputs bauen (P=1 Vergangenheit, S=1 neues Token)")
B, S, P = 1, 1, 1
dummy_embeds = torch.zeros(B, S, HIDDEN, dtype=torch.float32)
dummy_mask = torch.ones(B, P + S, dtype=torch.long)
dummy_pos = torch.tensor([[P]], dtype=torch.long)

# Pro Layer eigene Dummy-Form — 256 vs. 512 Head-Dim!
dummy_past = []
for (n_kv, hd) in KV_SHAPES:
    dummy_past.append(torch.zeros(B, n_kv, P, hd, dtype=torch.float32))
    dummy_past.append(torch.zeros(B, n_kv, P, hd, dtype=torch.float32))

input_names = ["inputs_embeds", "attention_mask", "position_ids"]
output_names = ["logits"]
dynamic_axes = {
    "inputs_embeds": {0: "batch", 1: "seq"},
    "attention_mask": {0: "batch", 1: "total"},
    "position_ids": {0: "batch", 1: "seq"},
    "logits": {0: "batch", 1: "seq"},
}
for i in range(N_CACHE):
    for kv in ("key", "value"):
        pn = f"past_key_values.{i}.{kv}"
        on = f"present.{i}.{kv}"
        input_names.append(pn)
        output_names.append(on)
        dynamic_axes[pn] = {0: "batch", 2: "past_seq"}
        dynamic_axes[on] = {0: "batch", 2: "total_seq"}

log("C) torch.onnx.export laeuft (dauert lange, kein Fortschrittsbalken)")
with torch.no_grad():
    torch.onnx.export(
        wrapper,
        (dummy_embeds, dummy_mask, dummy_pos, *dummy_past),
        str(FP32_ONNX),
        input_names=input_names,
        output_names=output_names,
        dynamic_axes=dynamic_axes,
        opset_version=17,
        do_constant_folding=True,
        dynamo=False,          # MUSS letztes Keyword bleiben
    )
print("Export geschrieben.")

del model, lm, lm_head, wrapper, dummy_past
gc.collect()


# --------------------------------------------------- D) Konsolidierung
log("D) Konsolidierung zu EINER .onnx_data (finaler Name direkt, nie umbenennen!)")
m = onnx.load(str(FP32_ONNX), load_external_data=True)
onnx.save_model(
    m,
    str(FP32_ONNX),
    save_as_external_data=True,
    all_tensors_to_one_file=True,
    location=FP32_DATA,
    size_threshold=1024,
)
del m
gc.collect()

# Fragment-Dateien aufraeumen (sonst laeuft die Disk voll)
for f in OUT.iterdir():
    if f.name.startswith("onnx__") or f.name.startswith("_"):
        f.unlink()
os.system(f"df -h /root; ls -la {OUT}")

onnx.checker.check_model(str(FP32_ONNX))
print("fp32-Graph valide.")


# ----------------------------------------------------- E) q4f16-Quantisierung
log("E) MatMul4BitsQuantizer -> q4f16 (der eine ungetestete Schritt)")
from onnxruntime.quantization.matmul_4bits_quantizer import (
    MatMul4BitsQuantizer,
    DefaultWeightOnlyQuantConfig,
)

model_fp32 = onnx.load(str(FP32_ONNX), load_external_data=True)

cfg = DefaultWeightOnlyQuantConfig(
    block_size=32,          # Transformers.js-kompatibel
    is_symmetric=True,
    accuracy_level=4,       # int8-Compute
)
quant = MatMul4BitsQuantizer(model_fp32, algo_config=cfg)
quant.process()

onnx.save_model(
    quant.model.model,
    str(Q4_ONNX),
    save_as_external_data=True,
    all_tensors_to_one_file=True,
    location=Q4_DATA,
    size_threshold=1024,
)
print("q4f16 geschrieben.")

del model_fp32, quant
gc.collect()


# --------------------------------------------------------- F) Verifikation
log("F) Verifikation")
onnx.checker.check_model(str(Q4_ONNX))

size_mb = (OUT / Q4_DATA).stat().st_size / 1e6
print(f"{Q4_DATA}: {size_mb:.0f} MB")
if size_mb > 3500:
    print("!! WARNUNG: >3.5 GB — Browser-ArrayBuffer-Limit gefaehrdet.")
else:
    print("OK: Groesse im browsertauglichen Bereich.")

import onnxruntime as ort

sess = ort.InferenceSession(str(Q4_ONNX), providers=["CPUExecutionProvider"])
print("Session laedt. Inputs:", len(sess.get_inputs()), "Outputs:", len(sess.get_outputs()))

log("FERTIG. Naechster Schritt: Tokenizer + config.json daneben legen,")
print("config.json braucht den transformers.js_config-Block mit")
print('  "use_external_data_format": true')
print("sonst wird die .onnx_data im Browser nie angefragt.")