The reasoning engine starves every other request while one long prompt
prefills. Measured: 262 tok/s contended prefill, ~250s for a 65k prompt,
with Avg generation throughput 0.0-0.1 tok/s at Running: 2-4 reqs.
Root cause is NOT what the issue first diagnosed. The issue proposed
--max-num-partial-prefills / --max-long-partial-prefills, which are V0-only
knobs. Verified in the live container: vLLM 0.11.0 ships V1 only (vllm/core/
is absent, so there is no V0 to fall back to), and EngineArgs.
_is_v1_supported_oracle returns False whenever either flag differs from its
default, calling _raise_or_fallback("Concurrent Partial Prefill") ->
"Falling back to V0" -> AssertionError. With strategy Recreate that is an
outage, not a slow rollout. Both flags are therefore NOT set, and a comment
records why so they are not reintroduced.
V1 needs no such flag: it has no prefill/decode phase at all. Every step it
walks self.running and hands out tokens from one shared budget
(vllm/v1/core/sched/scheduler.py). The real defect is that the per-request
cap is disabled by default (long_prefill_token_threshold=0), so one long
prompt takes the WHOLE budget, token_budget hits 0, the `while ... and
token_budget > 0` loop exits, and every other request is skipped.
The cap is applied before the budget clamp, so threshold >= budget never
binds. Replaying the loop: 8192/8192 serves 1 request per step (still
starved); 8192/2048 serves 4, and two long prompts advance together, which
is the issue's actual acceptance criterion.
--max-num-batched-tokens=8192 (was unset -> default 2048)
--long-prefill-token-threshold=2048 (was unset -> 0, cap disabled)
--max-num-seqs=5, --gpu-memory-utilization=0.95
The ~166k KV figure is marked as a projection; the live reading is
num_gpu_blocks=9527 x block_size=16 = 152,432 tokens and is authoritative
post-boot. A preemption alert covers the accept-and-watch risk.
Observability for the companion admission gate (homelab-frontend#64, now
deployed at 64f2bfc1):
- four dashboard panels: parked, rejected by reason, engine state stale,
and gateway-view vs engine-truth. Each carries a description, because
an unexplained blank panel reads as a broken dashboard.
- alerts on observer_degraded and engine_state_stale_total, both with an
absent() guard: the gate fails OPEN, so a missing or renamed metric must
alert rather than look healthy. Verified live that the guards do not
false-fire (both series present, 3 each).
- LLMPreemptions and LLMPrefillStarvation: the acceptance metric and the
regression direction this change introduces.
- max(vllm_requests_running), not sum(): one engine-global gauge is read
by all three gateway replicas, so sum() read 3x and could never agree.
Also corrects comments that were factually wrong. An earlier revision of
this branch claimed llm_ttft_seconds/llm_itl_seconds/llm_tokens_total were
dead code; homelab-frontend#63 wired them, verified live as 3 series each.
request_duration_seconds is the genuinely dead one (0 samples, zero
non-test call sites) and is now marked as such instead of being listed as
working.
Closes#51
Previous comments described llm_ttft_seconds / llm_itl_seconds /
llm_tokens_total as "streaming-only". That was wrong, and optimistic.
They are dead code. internal/proxy/llm_metrics.go defines
LLMMetricsCapture, which is the only thing that calls RecordTTFT /
RecordITL / RecordTokens — but NewLLMMetricsCapture() has zero call sites
anywhere in the repo, verified by grep. The exporter skips keys with no
samples, so those three series render as HELP-only stubs and can never be
graphed, streaming or not.
No behaviour change here; only the comments and the dashboard generator's
header are corrected so the next person does not build on a false
assumption. The gateway latency panels are left in place so the gap stays
visible — they populate unchanged once the capture wrapper is wired into
the streaming proxy path.
Series that DO work from the gateway, confirmed live with 3 series (one
per replica): queue_full_requests_total, warmup_requests_delayed_total,
vllm_prefix_cache_hit_tokens_total.
New Grafana dashboard (llm-observability) with 27 panels across 5 rows:
TTFT, inter-token latency, batching/scheduling, KV cache, throughput.
Both metric sources are shown side by side because they answer different
questions and neither is complete alone:
vLLM 0.11.0 (reasoning) — engine truth: kv_cache_usage_perc,
num_requests_running/waiting, num_preemptions_total, prefill vs
decode time, prefix cache, per-request histogram quantiles.
TEI 1.8.2 (embeddings, reranker) — te_batch_next_size, te_queue_size,
te_embed_duration, te_embed_queue_duration.
api-gateway — the ONLY per-request latency source covering Ollama
models (ornith:35b, qwen2.5:3b-instruct), which expose no metrics.
It is derived from the SSE byte stream, so it is streaming-only and
counts SSE chunks rather than true tokens. vLLM's own histograms are
authoritative for reasoning.
All 44 panel queries were executed against live Prometheus before commit —
all return success. vLLM/TEI panels have live series; gateway panels show
no data until the netpol fix in homelab-frontend PR #62 merges.
Also fixes a dead alert. alerts/cag-kv-cache-rules.yaml derived
vllm:prefix_cache_hit_rate from vllm:cache_query_hit_total and
vllm:cache_query_miss_total. Those metric names do not exist in vLLM
0.11.0 — the real series are vllm:prefix_cache_hits_total and
vllm:prefix_cache_queries_total. The recording rule produced no data, so
VLLMPrefixCacheHitRateLow could never fire.
Not included: real GPU hardware metrics (SM utilisation, framebuffer,
power, temperature). Only nvidia-device-plugin is deployed, which exposes
none of them; DCGM exporter is required. vllm:kv_cache_usage_perc and
preemption rate are the only GPU-pressure proxies available today.
Ingress api/api now backs onto api-gateway:8080; the kong Application, its
Helm values, plugins and llm-routes are removed. Gateway image v0.0.0 is in
the Forgejo registry and the pull secret is in the api namespace.
rock
Story Crater Bot