File size: 2,410 Bytes
d548d67
14690b9
 
 
 
 
 
 
 
 
d548d67
 
 
 
14690b9
d548d67
 
 
 
14690b9
d548d67
e6a8d26
d548d67
 
 
14690b9
d548d67
 
 
 
e6a8d26
14690b9
e6a8d26
 
 
 
 
 
 
 
 
d548d67
 
14690b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d548d67
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
#!/usr/bin/env python3
"""Export joemaddalone/ffmpeg-nl2cmd-0.8B to GGUF Q8_0 WITHOUT the MTP tail
(--no-mtp), verify the tensor map, and upload to the model repo.

The previous export kept the MTP speculative-decode block as blk.24.*, which
ollama's llama.cpp rejects ('blk.24.attn_norm.weight not found').
"""
import os
import subprocess
import sys

REPO = "joemaddalone/ffmpeg-nl2cmd-0.8B"
WORK = "/workspace"


def sh(cmd, **kw):
    print("+", cmd, flush=True)
    return subprocess.run(cmd, shell=True, **kw)


sh("pip install -q --no-cache-dir gguf numpy 'torch<3' transformers sentencepiece protobuf || true")
sh("apt-get update -qq && apt-get install -y -qq git >/dev/null 2>&1 || true")
if not os.path.isdir(f"{WORK}/llama.cpp"):
    sh(f"git clone --depth 1 https://github.com/ggml-org/llama.cpp {WORK}/llama.cpp")

from huggingface_hub import snapshot_download, HfApi  # noqa: E402
local = snapshot_download(REPO)
print("model at", local, flush=True)

out = f"{WORK}/ffmpeg-nl2cmd-0.8B-q8_0.gguf"
ok = False
for extra in (["--no-mtp"], ["--no-mtp", "--disable-vision"]):
    r = sh(f"python {WORK}/llama.cpp/convert_hf_to_gguf.py {local} --outfile {out} --outtype q8_0 " +
           " ".join(extra))
    if r.returncode == 0 and os.path.exists(out):
        print("converted with extra args:", extra, flush=True)
        ok = True
        break
    print("conversion failed with extra args:", extra, flush=True)
if not ok:
    print("CONVERT_FAILED (all fallbacks)", flush=True)
    sys.exit(1)

# --- verify tensor map: no MTP block (blk.24), last real block complete ---
from gguf import GGUFReader  # noqa: E402
reader = GGUFReader(out)
names = [t.name for t in reader.tensors]
blocks = sorted({int(n.split(".")[1]) for n in names if n.startswith("blk.")})
mtp = [n for n in names if n.startswith("blk.24.")]
print(f"tensors: {len(names)} | blocks: {blocks[0]}..{blocks[-1]}", flush=True)
if mtp:
    print("MTP tensors still present:", mtp[:5], flush=True)
    sys.exit(1)
if f"blk.{blocks[-1]}.attn_norm.weight" not in names:
    print("last block incomplete!", flush=True)
    sys.exit(1)
print(f"VERIFIED: no MTP block, blk.{blocks[-1]} complete", flush=True)

print("gguf size:", os.path.getsize(out), flush=True)
api = HfApi()
api.upload_file(path_or_fileobj=out, path_in_repo="ffmpeg-nl2cmd-0.8B-q8_0.gguf",
                repo_id=REPO, repo_type="model")
print("UPLOADED", flush=True)