Isaac Quarenta commited on
Commit
dccbaf4
·
1 Parent(s): 13413b9

fix(kaggle-cpu): resolver binario via nightly ggml-org (latest versionada nao tem assets)

Browse files
Files changed (1) hide show
  1. kaggle_backend.py +34 -9
kaggle_backend.py CHANGED
@@ -82,21 +82,46 @@ def _llama_server_bin(workdir):
82
  p = os.path.join(root, "llama-server")
83
  os.chmod(p, os.stat(p).st_mode | stat.S_IEXEC)
84
  return p
85
- api = "https://api.github.com/repos/ggerganov/llama.cpp/releases/latest"
86
- req = urllib.request.Request(api, headers={"Accept": "application/vnd.github+json"})
87
- _gh = os.getenv("GH_TOKEN", "").strip()
88
- if _gh:
89
- req.add_header("Authorization", f"Bearer {_gh}")
 
 
 
 
 
 
 
 
90
  print("[kaggle-backend:cpu] a descobrir release llama.cpp ...", flush=True)
91
- with urllib.request.urlopen(req, timeout=60) as r:
92
- rel = _j.loads(r.read().decode())
 
 
 
 
 
 
 
 
 
 
 
93
  asset = None
94
- for a in rel.get("assets", []):
95
  if a.get("name", "").endswith("bin-ubuntu-x64.zip"):
96
  asset = a
97
  break
 
 
 
 
 
 
98
  if asset is None:
99
- raise RuntimeError("release sem bin-ubuntu-x64.zip")
100
  zpath = os.path.join(workdir, asset["name"])
101
  if not os.path.exists(zpath):
102
  print(f"[kaggle-backend:cpu] a baixar {asset['name']} ...", flush=True)
 
82
  p = os.path.join(root, "llama-server")
83
  os.chmod(p, os.stat(p).st_mode | stat.S_IEXEC)
84
  return p
85
+ import re as _re
86
+
87
+ def _gh_json(url):
88
+ req = urllib.request.Request(url, headers={"Accept": "application/vnd.github+json"})
89
+ _gh = os.getenv("GH_TOKEN", "").strip()
90
+ if _gh:
91
+ req.add_header("Authorization", f"Bearer {_gh}")
92
+ with urllib.request.urlopen(req, timeout=60) as r:
93
+ return _j.loads(r.read().decode())
94
+
95
+ # O repo mudou para ggml-org e as releases versionadas (v0.6.0 etc.)
96
+ # já NÃO trazem binários — só as nightlies (tags bNNNN). A latest
97
+ # indica a nightly no corpo ("Nightly build: [b11429](.../tag/b11429)").
98
  print("[kaggle-backend:cpu] a descobrir release llama.cpp ...", flush=True)
99
+ rel = _gh_json("https://api.github.com/repos/ggml-org/llama.cpp/releases/latest")
100
+ assets = []
101
+ mtag = _re.search(r"releases/tag/(b\d+)", rel.get("body", "") or "")
102
+ if mtag:
103
+ try:
104
+ nrel = _gh_json(
105
+ "https://api.github.com/repos/ggml-org/llama.cpp/releases/tags/" + mtag.group(1))
106
+ assets = nrel.get("assets", [])
107
+ print(f"[kaggle-backend:cpu] nightly {mtag.group(1)}: {len(assets)} assets", flush=True)
108
+ except Exception as e:
109
+ print(f"[kaggle-backend:cpu] nightly falhou ({e}), a usar latest", flush=True)
110
+ if not assets:
111
+ assets = rel.get("assets", [])
112
  asset = None
113
+ for a in assets:
114
  if a.get("name", "").endswith("bin-ubuntu-x64.zip"):
115
  asset = a
116
  break
117
+ if asset is None: # fallback: qualquer zip ubuntu-x64
118
+ for a in assets:
119
+ nm = a.get("name", "")
120
+ if "ubuntu-x64" in nm and nm.endswith(".zip"):
121
+ asset = a
122
+ break
123
  if asset is None:
124
+ raise RuntimeError("release sem bin-ubuntu-x64.zip (nem fallback ubuntu-x64)")
125
  zpath = os.path.join(workdir, asset["name"])
126
  if not os.path.exists(zpath):
127
  print(f"[kaggle-backend:cpu] a baixar {asset['name']} ...", flush=True)