95 waiters all tokenizing 2M-token docs through the 32-thread default
executor saturated the GIL: the rich render thread and the event loop
starved, so the bar froze and jumped (and the gate probe went blind).
Truncation now runs on a dedicated 8-thread pool; the remaining
workers queue and the loop/renderer stay responsive.
eta formats as 45s / 7m15s / 2h35m40s / 6d03h12m as it grows.
Co-Authored-By: Claude <noreply@anthropic.com>
The 5s probe fetched /metrics via asyncio.to_thread, which shares the
DEFAULT executor with second-long truncation tokenizations -- 96 of
those queue-jumped the probe and the state machine never ticked (gate
frozen at 1 while results flowed, ETA 6h). Probes now run on a
dedicated single-thread executor, and after 3 consecutive fetch
failures the gate stops asking for /metrics entirely (this endpoint
404s; pure demand mode from then on).
Verified under a choked default executor: gate ticks 1->2 on schedule
and metrics_dead engages after 3 failures.
Co-Authored-By: Claude <noreply@anthropic.com>
96 worker threads racing transformers 5.x lazy imports on the FIRST
_get_tokenizer call raised ImportError and degraded that whole first
batch to the char approximation (the old single-threaded path never
raced). First load now holds a threading.Lock; the transformers
'>model_max_length' logging is silenced inside truncation (counting a
2M-token doc before trimming it is the point), and the per-sample
degradation print becomes a one-shot warning.
Co-Authored-By: Claude <noreply@anthropic.com>
The counter was taken just past the GLOBAL semaphore (lifted to 96 in
auto mode so the gate is the sole limiter) but BEFORE the pool gate --
so 94 gate-queued workers counted as in-flight. The gate now pushes
its actually-admitted count and the bar shows 'admitted/held'
(e.g. in-flight 2/96 = 2 really hitting the server, 94 queued on the
gate). Verified: admitted never exceeds the gate limit.
Co-Authored-By: Claude <noreply@anthropic.com>
Replaces the +1/5s linear ramp: start at 1, double while measured
completions/s keeps improving (>10% over the previous level); the
first plateau opens a bisect [last_good, bad] that narrows to the
knee, then holds steady. Failures still cut x0.7 instantly and
restart probing from the shrunken level; zero completions = hold.
Judging a level needs max(MIN_OK, level) completions -- a
2-completion rate estimate at level 8 is quantization noise (caught
by simulation converging to 1 on a capacity-8 endpoint).
Simulated against throughput curves min(level, capacity):
capacity 8 -> 1,2,4,8,16 | bisect 12,10,9 -> steady 8
capacity 16 -> 1,2,4,8,16,32 | bisect ... -> steady 16
capacity 4 -> 1,2,4,8 | bisect 6,5 -> steady 4
Co-Authored-By: Claude <noreply@anthropic.com>
Ramping on demand alone was dangerous with slow endpoints: if the 2
in-flight longbench_v2 requests hang, 'no failures + waiters queued'
kept adding +1 every 5s all the way to 96 -- piling prefills onto a
server that had not answered anything. Both ramp paths (demand-driven
and /metrics) now require at least one SUCCESSFUL completion in the
probe interval; hangs hold the gate until read-timeouts fire and the
x0.7 backoff takes over.
Unit-verified: hang 4 probe intervals with 20 waiters -> limit stays 2;
one success -> +1; one failure -> x0.7.
Co-Authored-By: Claude <noreply@anthropic.com>
assemble() runs the max_input_tokens truncation tokenizer inline in the
coroutine; longbench_v2's 2M-token docs take seconds of CPU per encode,
and each one BLOCKED the whole loop -- heartbeat frozen, gate probes
dead, zero HTTP while the process sat at 100% single-core. encode now
runs in a worker thread (loop stays live, encodes parallelize).
Co-Authored-By: Claude <noreply@anthropic.com>
The auto-stream trigger only looked at max_tokens (output budget);
longbench_v2 sends ~128k-token INPUTS with a small 8k output budget and
hit the gateway's whole-request buffering: 20-minute read timeouts
through all 6 retries. es survives the same bench because it streams
by default. Payloads over ~300k chars (~90k+ tokens) now take the
stream-aggregate path; verified live: a 350k-char prompt returns in
20s instead of hanging.
Co-Authored-By: Claude <noreply@anthropic.com>
Full traceback finally caught it: adapter._parse ALWAYS ran the
text-protocol tool-call fallback, even for requests with NO tools. On
humaneval, model code like regex-matched as a
'call', ast.literal_eval turned the literal into an Ellipsis
(no exception -- it's a legal literal), and json.dumps(args) died
mid-generation, killing the benchmark.
Two layers:
- the fallback now only runs when the request actually carried tools
(also stops polluting plain predictions with phantom calls, and the
SyntaxWarning spam from ast.parse-ing model code disappears)
- json.dumps(args, default=str) as belt-and-braces for the
text-tools path where an Ellipsis arg now stringifies
Reproduced the exact crash input as a unit case: no-tools code reply
yields 0 tool_calls; in text mode serializes
{'key': 'Ellipsis'} without raising; normal fc calls unchanged.
Co-Authored-By: Claude <noreply@anthropic.com>
User feedback: the out-dir report-reuse layer was one concept too many.
Now each checkpoint line carries {key, ts, pred, score}:
- --resume restores predictions AND their scores; when every sample's
cached score matches the scoring-setup fingerprint (recipe/extract/
scorers/judge), the report is replayed with NO scorer, extractor or
docker container touching anything
- fingerprint mismatch (recipe or judge changed) -> automatic re-eval
and backfill of the fresh scores
- no --resume -> nothing read, nothing written (full fresh run)
- --rescore = ignore cached scores, re-evaluate, refresh the cache
- aggregation always recomputed from cached per-sample scores (cheap,
survives aggregator changes without invalidating)
- legacy checkpoints without a score field backfill on first evaluation
Removed: the out-dir report-reuse block (superseded; also the source of
the UnboundLocalError path).
Co-Authored-By: Claude <noreply@anthropic.com>
--auto-concurrency wraps a single endpoint as a 1-member pool; the
spec parser rejected it with 'needs a {start..end} port range' and
killed every benchmark. Range-less comma segments are now plain
members; only a fully empty endpoint list errors.
Co-Authored-By: Claude <noreply@anthropic.com>
- new flag wraps even a single endpoint as a one-member pool so the
per-endpoint AdaptiveGate takes over; --concurrency becomes the gate's
STARTING point (global semaphore lifted to the gate ceiling of 96)
- demand-driven AIMD fallback for endpoints without /metrics (404,
gateway-stripped, non-sglang): ramp +1 while callers wait on acquire
and the interval is failure-free; the old code early-returned on
fetch errors and never adapted at all. Real GLM endpoint verified:
/metrics is 404, so this fallback is the live path there
- probe fetch moved to a thread: a blocked urlopen parked the whole
event loop for its 4s timeout
- current limit surfaces on the progress bar as a magenta 'gate N'
field (pushed on every change)
Co-Authored-By: Claude <noreply@anthropic.com>
Scoring progress: docker-exec benches (humaneval etc.) score for
minutes with zero feedback -- the bar sat at 'generating 100%' and
looked hung. evaluate() now takes on_scored(i, n) (atomic counter,
fires from worker threads), run_eval passes it through, and the CLI
shows 'scoring 42/164' on the bar + milestone log lines every 10%
(also fixes the phase match: 'scoring' never matched the capitalized
'Scoring predictions...' status message, so the bar never even
switched its label).
PooledAdapter:
- one release per acquire: the exception path released True (inner
finally) AND False (except handler), double-decrementing _inflight
(over-admission) and applying the x0.7 backoff twice
- AdaptiveGate: rebuild the Condition + probe task when the event loop
changes -- pools are cached across benchmarks and the CLI runs
asyncio.run() per bench/repeat; a loop-bound Condition from a closed
loop raises 'bound to a different event loop' under contention
Co-Authored-By: Claude <noreply@anthropic.com>
- perf_stats aggregator lives in eval/, not model/: the import failed
silently and EVERY perf column was empty (not just ttft). Now warns
on stderr instead of swallowing.
- repeats > 1 get their own checkpoint key (:rep2, :rep3, ...): repeat 2
previously restored repeat 1's predictions and finished instantly with
identical scores. rep1 keeps the legacy key (existing checkpoints still
resume).
- repeats summary: report the MEAN score and aggregate time/tokens over
ALL runs (was: last run only).
- README: six-benchmark command as the primary example.
Co-Authored-By: Claude <noreply@anthropic.com>
Was: raw load_raw_records() hit the hub on EVERY run -- offline machines
stalled in 5x HF retries then silently degraded to 0-shot (changing the
benchmark's default contract, e.g. gsm8k 4-shot). Now the few-shot split
is a proper Dataset entry: first use downloads+cache, every later run is
a pure cache hit (verified: second run with HF_HUB_OFFLINE=1 loads
4-shot from cache, zero network lines, acc unchanged).
Co-Authored-By: Claude <noreply@anthropic.com>
Explicit key flows only into request headers (adapter attribute / pool
member), so it cannot leak into the spec string, EvalReport, or logs --
verified by scanning a report produced with a sentinel key. Two-key
setups run twice with different --api-key, or use per-host env vars.
Co-Authored-By: Claude <noreply@anthropic.com>
- --judge now accepts a bare model name paired with --judge-api-url,
mirroring --model/--api-url; full legacy specs keep working
(_compose_judge_spec, verified: name+url -> spec, spec passthrough,
empty -> None; end-to-end on simple_qa with a live judge endpoint)
- mock adapter spellings: mock-boxed / mock-oracle / mock-fc preferred,
colon forms still accepted; bare 'mock' stays echo
- fix mock adapter singleton mode pollution: resolve_adapter memoizes
one instance, so mock-boxed then mock in one process leaked the
boxed mode into the echo run -- each mock spec now builds a fresh
instance
- README: mock-boxed in examples, judge flags row updated
Co-Authored-By: Claude <noreply@anthropic.com>
