ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into feat/gefenx-optimizer

Browse source
This commit is contained in:
thad0ctor 2026-07-10 08:53:21 -07:00 • committed by GitHub
commit a03d85ba69
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 120 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

44
.github/workflows/studio-inference-smoke.yml vendored
View file

@ -444,6 +444,8 @@ jobs:
python - <<'PY'
import json
import os
import time
import urllib.error
import urllib.request
BASE = os.environ["BASE_URL"]
@ -464,8 +466,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
def post_sse(path, body, *, timeout = 600):
"""POST a streaming request and accumulate the assistant
@ -938,6 +956,8 @@ jobs:
import base64
import json
import os
import time
import urllib.error
import urllib.request
from openai import OpenAI
from anthropic import Anthropic
@ -956,8 +976,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
# ── 1. response_format = json_object (JSON mode) ─────────────
# llama.cpp's HTTP server supports OpenAI-compatible JSON

44
.github/workflows/studio-mac-inference-smoke.yml vendored
View file

@ -430,6 +430,8 @@ jobs:
python - <<'PY'
import json
import os
import time
import urllib.error
import urllib.request
BASE = os.environ["BASE_URL"]
@ -450,8 +452,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
def post_sse(path, body, *, timeout = 600):
"""POST a streaming request and accumulate the assistant
@ -825,6 +843,8 @@ jobs:
import base64
import json
import os
import time
import urllib.error
import urllib.request
from openai import OpenAI
from anthropic import Anthropic
@ -848,8 +868,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
# ── 1. response_format = json_object (JSON mode) ─────────────
# llama.cpp's HTTP server supports OpenAI-compatible JSON

44
.github/workflows/studio-windows-inference-smoke.yml vendored
View file

@ -634,6 +634,8 @@ jobs:
python - <<'PY'
import json
import os
import time
import urllib.error
import urllib.request
BASE = os.environ["BASE_URL"]
@ -656,8 +658,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
def post_sse(path, body, *, timeout = 600):
body = {**body, "stream": True}
@ -1063,6 +1081,8 @@ jobs:
import base64
import json
import os
import time
import urllib.error
import urllib.request
from openai import OpenAI
from anthropic import Anthropic
@ -1082,8 +1102,24 @@ jobs:
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(req, timeout = timeout) as resp:
return resp.status, json.loads(resp.read().decode())
# Shared CI runners stall sporadically, so retry transport-level
# failures only; HTTP status errors surface immediately. Bounded
# to fit the job's timeout-minutes: short probes get 3 full
# attempts, long probes one retry capped at 300s (a healthy
# server answers a retry quickly; a stalled one never does).
attempts = 3 if timeout <= 300 else 2
for attempt in range(attempts):
try:
t = timeout if attempt == 0 else min(timeout, 300)
with urllib.request.urlopen(req, timeout = t) as resp:
return resp.status, json.loads(resp.read().decode())
except urllib.error.HTTPError:
raise
except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
if attempt == attempts - 1:
raise
print(f"[retry] {path}: {exc!r}", flush = True)
time.sleep(15)
# ── 1. response_format = json_object (JSON mode) ─────────────
status, data = post("/v1/chat/completions", {