07_export_quant_c_v2.py

#!/usr/bin/env python
"""
07_export_quant_c_v2.py — WEG C, KORRIGIERTE fp16-Konvertierung.

Unterschied zu 03_export_quant_c.py:
  - D.5 haelt RMSNorm/Softcap in fp32 (Default-op_block_list) statt alles fp16 zu
    erzwingen. Dafuer wird VOR der Konvertierung dateibasierte Shape-Inference
    gefahren (infer_shapes_path, vertraegt >2GB), damit convert_float_to_float16
    die Cast-Bruecken an den fp32/fp16-Grenzen korrekt setzt.
  - KEIN fix_fp16_edges mehr (wir WOLLEN die fp32-Inseln — sie verhindern den
    fp16-Ueberlauf in der RMSNorm-Reduktion auf echten WebGPU-Kerneln).

Grund: op_block_list=[] erzwang fp16 ueberall -> ReduceMean-Summe (mean(x^2) ueber
2560 Dims, x~50 nach Gemma-Normalizer) sprengt fp16-Max (65504) auf WebGPU -> NaN.
ORT-CPU rechnet intern fp32 und verzeiht das (deshalb war der CPU-Test kohaerent).

Start IMMER mit nohup:
  nohup python 07_export_quant_c_v2.py > export_v2.log 2>&1 &
"""
import gc, os
from pathlib import Path
import torch, onnx

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

MODEL_ID = "/root/gemma4-bund-merged"          # LOKAL (kein Re-Download)
STOCK    = "onnx-community/gemma-4-E4B-it-ONNX"
OUT      = Path("/root/train/gemma4-bund-final-v2")
ONNX_DIR = OUT / "onnx"; ONNX_DIR.mkdir(parents=True, exist_ok=True)

FP32      = ONNX_DIR / "decoder_model_merged.onnx"
FP32_DATA = "decoder_model_merged.onnx_data"
FP32_SI   = ONNX_DIR / "decoder_model_merged_si.onnx"        # shape-inferred
FP32_SI_DATA = "decoder_model_merged_si.onnx_data"
FP16      = ONNX_DIR / "decoder_model_merged_fp16.onnx"
FP16_DATA = "decoder_model_merged_fp16.onnx_data"
Q4        = ONNX_DIR / "decoder_model_merged_q4f16.onnx"
Q4_DATA   = "decoder_model_merged_q4f16.onnx_data"

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

# ------------------------------------------------------------------ A) laden
log("A) Modell laden (fp32)")
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()
lm = model.model.language_model
lm_head = model.lm_head if hasattr(model, "lm_head") else model.get_output_embeddings()
print("hidden_size:", lm.config.hidden_size)

# ------------------------------------------------- B) Cache-Geometrie messen
log("B) Trockenlauf")
with torch.no_grad():
    ids = torch.tensor([[1, 2, 3, 4]])
    emb = lm.get_input_embeddings()(ids)
    ple = lm.get_per_layer_inputs(ids, emb)
    print("per_layer_inputs Shape:", tuple(ple.shape), "(erwartet: 1,4,42,256)")
    probe = lm(inputs_embeds=emb, per_layer_inputs=ple, use_cache=True, return_dict=True)
pkv = probe.past_key_values
N_CACHE = len(pkv.layers)
print("n_cache_layers:", N_CACHE, "(erwartet: 24)")
KV_SHAPES = [(int(pkv.layers[i].keys.shape[1]), int(pkv.layers[i].keys.shape[3])) for i in range(N_CACHE)]
print("head_dims:", sorted({s[1] for s in KV_SHAPES}), "(erwartet: [256, 512])")
del probe, pkv, emb, ple, ids; gc.collect()

# ---------------------------------------------------------------- C) Wrapper
class DecoderWrapper(torch.nn.Module):
    def __init__(self, lm, lm_head, n_cache):
        super().__init__(); self.lm, self.lm_head, self.n_cache = lm, lm_head, n_cache
    def forward(self, inputs_embeds, per_layer_inputs, 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, per_layer_inputs=per_layer_inputs,
                      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")
B, S, P = 1, 1, 1
with torch.no_grad():
    d_ids = torch.tensor([[42]], dtype=torch.long)
    d_emb = lm.get_input_embeddings()(d_ids).detach()
    d_ple = lm.get_per_layer_inputs(d_ids, d_emb).detach()
d_mask = torch.ones(B, P + S, dtype=torch.long)
d_pos = torch.tensor([[P]], dtype=torch.long)
d_past = []
for (n_kv, hd) in KV_SHAPES:
    d_past += [torch.zeros(B, n_kv, P, hd), torch.zeros(B, n_kv, P, hd)]

input_names = ["inputs_embeds", "per_layer_inputs", "attention_mask", "position_ids"]
output_names = ["logits"]
dyn = {"inputs_embeds":{0:"batch",1:"seq"}, "per_layer_inputs":{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, on = f"past_key_values.{i}.{kv}", f"present.{i}.{kv}"
        input_names.append(pn); output_names.append(on)
        dyn[pn] = {0:"batch",2:"past_seq"}; dyn[on] = {0:"batch",2:"total_seq"}

log("C) torch.onnx.export — LANGE STILLE IST NORMAL")
with torch.no_grad():
    torch.onnx.export(wrapper, (d_emb, d_ple, d_mask, d_pos, *d_past), str(FP32),
        input_names=input_names, output_names=output_names, dynamic_axes=dyn,
        opset_version=17, do_constant_folding=True, dynamo=False)
print("Export geschrieben.")
del model, lm, lm_head, wrapper, d_past, d_emb, d_ple; gc.collect()

# --------------------------------------------------------- D) Konsolidierung
log("D) Konsolidierung")
m = onnx.load(str(FP32), load_external_data=True)
onnx.save_model(m, str(FP32), save_as_external_data=True, all_tensors_to_one_file=True,
                location=FP32_DATA, size_threshold=1024)
del m; gc.collect()
for f in ONNX_DIR.iterdir():
    if f.name.startswith("onnx__") or f.name.startswith("lm.") or f.name.startswith("_"):
        f.unlink()
try:
    onnx.checker.check_model(str(FP32)); print("fp32 valide.")
except Exception as e:
    print("checker uebersprungen (>2GB):", type(e).__name__)

# ---------------------------- D.5) KORRIGIERT: Shape-Inference + fp16, RMSNorm bleibt fp32
log("D.5) Shape-Inference (dateibasiert, >2GB-tauglich)")
from onnx import shape_inference
shape_inference.infer_shapes_path(str(FP32), str(FP32_SI))
print("shape-inferred geschrieben.")
FP32.unlink(missing_ok=True); (ONNX_DIR / FP32_DATA).unlink(missing_ok=True)

log("D.5) fp32 -> fp16 (Default-op_block_list: RMSNorm/Softcap bleiben fp32)")
from onnxconverter_common import float16
m32 = onnx.load(str(FP32_SI), load_external_data=True)
# KEIN op_block_list=[] -> Default-Liste haelt numerisch heikle Ops in fp32.
# disable_shape_infer=True ist ok, weil FP32_SI bereits value_info traegt.
m16 = float16.convert_float_to_float16(m32, keep_io_types=False, disable_shape_infer=True)
onnx.save_model(m16, str(FP16), save_as_external_data=True, all_tensors_to_one_file=True,
                location=FP16_DATA, size_threshold=1024)
del m32, m16; gc.collect()
FP32_SI.unlink(missing_ok=True); (ONNX_DIR / FP32_SI_DATA).unlink(missing_ok=True)
print("fp16 geschrieben (mit fp32-Inseln).")

# Verifikation: laedt + RMSNorm-Reduktion fp32?
import onnxruntime as ort
try:
    so = ort.SessionOptions(); so.intra_op_num_threads = 4
    ort.InferenceSession(str(FP16), sess_options=so, providers=["CPUExecutionProvider"])
    print("fp16 LAEDT in ORT.")
except Exception as e:
    print("!! fp16 LAEDT NICHT:", str(e).split(chr(10))[0][:120])
rm_fp32 = 0
mm = onnx.load(str(FP16), load_external_data=False)
vi = {v.name: v.type.tensor_type.elem_type for v in mm.graph.value_info}
for n in mm.graph.node:
    if n.op_type == "ReduceMean" and vi.get(n.output[0]) == 1:
        rm_fp32 += 1
print(f"ReduceMean in fp32: {rm_fp32} (erwartet >0)")
del mm; gc.collect()

# ------------------------------------------------------------------ E) q4f16
log("E) q4f16")
from onnxruntime.quantization.matmul_nbits_quantizer import (
    MatMulNBitsQuantizer as Q, DefaultWeightOnlyQuantConfig)
mf = onnx.load(str(FP16), load_external_data=True)
quant = Q(mf, algo_config=DefaultWeightOnlyQuantConfig(block_size=32, is_symmetric=True, accuracy_level=4))
quant.process()
qm = quant.model.model if hasattr(quant.model, "model") else quant.model
onnx.save_model(qm, str(Q4), save_as_external_data=True, all_tensors_to_one_file=True,
                location=Q4_DATA, size_threshold=1024)
del mf, quant, qm; gc.collect()
FP16.unlink(missing_ok=True); (ONNX_DIR / FP16_DATA).unlink(missing_ok=True)

# Verifikation nach Quantisierung
try:
    so = ort.SessionOptions(); so.intra_op_num_threads = 4
    s = ort.InferenceSession(str(Q4), sess_options=so, providers=["CPUExecutionProvider"])
    print("q4f16 LAEDT, Inputs total", len(s.get_inputs()))
except Exception as e:
    print("!! q4f16 LAEDT NICHT:", str(e).split(chr(10))[0][:120])

# ---------------------------------------------------- F) Stock-Embed + Config
log("F) Stock-Embed holen, Embed-Output auf fp16 casten, Config/Tokenizer schreiben")
from huggingface_hub import hf_hub_download
import shutil, json
from onnx import helper, TensorProto
for fn in ("onnx/embed_tokens_q4f16.onnx", "onnx/embed_tokens_q4f16.onnx_data"):
    p = hf_hub_download(STOCK, fn, local_dir="/root/stock-embed")
    shutil.copy(p, ONNX_DIR / Path(fn).name); print("geholt:", fn)

# Embed-Outputs fp32 -> fp16 casten (sonst dtype-Mismatch zum fp16-Decoder)
emb_path = ONNX_DIR / "embed_tokens_q4f16.onnx"
em = onnx.load(str(emb_path), load_external_data=False)
tgts = [o.name for o in em.graph.output if o.type.tensor_type.elem_type == TensorProto.FLOAT]
prod = {out: (nd, i) for nd in em.graph.node for i, out in enumerate(nd.output) if out in tgts}
for name in tgts:
    nd, idx = prod[name]; pre = name + "_fp32"; nd.output[idx] = pre
    em.graph.node.append(helper.make_node("Cast", [pre], [name], to=TensorProto.FLOAT16, name=name+"/CastToFp16"))
    for o in em.graph.output:
        if o.name == name: o.type.tensor_type.elem_type = TensorProto.FLOAT16
onnx.save(em, str(emb_path))
print("Embed-Outputs auf fp16 gecastet:", tgts)

tok.save_pretrained(str(OUT))
from transformers import AutoConfig
c = AutoConfig.from_pretrained(MODEL_ID); c.save_pretrained(str(OUT))
cp = OUT / "config.json"; cfg = json.load(open(cp))
cfg["transformers.js_config"] = {
    "dtype": "q4f16",
    "use_external_data_format": {
        "decoder_model_merged_q4f16.onnx": 2,   # 2 Chunks: .onnx_data + .onnx_data_1
        "embed_tokens_q4f16.onnx": True,
    },
    "kv_cache_dtype": "float16",
}
json.dump(cfg, open(cp, "w"), indent=2)

# chat_template ins tokenizer_config.json einbetten (Transformers.js liest es nur von dort)
tcp = OUT / "tokenizer_config.json"
jinja = OUT / "chat_template.jinja"
if jinja.exists():
    tc = json.load(open(tcp)); tc["chat_template"] = jinja.read_text(encoding="utf-8")
    json.dump(tc, open(tcp, "w"), ensure_ascii=False, indent=2)
    print("chat_template in tokenizer_config.json eingebettet.")

log("FERTIG bis E/F. NAECHSTE SCHRITTE:")
print("  1) Reshard:  python 04_reshard.py   (REPO/BASE ggf. anpassen, Quelle = dieser Q4)")
print("  2) Upload nach gfp78/gemma4-bund-onnx-v2 (oder v1 ueberschreiben)")
print("  3) Bundesrechner: 3 Decoder-Dateien + embed + config + tokenizer_config ziehen, Browser-Test")
print("  !! JETZT SOFORT die q4f16-Dateien sichern — /root ist fluechtig.")