Router: new /metrics/scatter endpoint returns individual data points
(prompt_tokens, inference_ms, model, agent, reason, stream)
for scatter visualization.
Dashboard: new panel showing latency vs prompt size by model.
- Log-scale X axis (prompt tokens) with model color coding
- Dropdown to filter by individual model or view all
- Hover tooltips with details per point
- Auto-refresh every 30s
Enables direct observation of context-length vs latency
relationship — validates routing tier decisions.
Router now buffers streaming response chunks to extract timings
(prompt_n, predicted_n, predicted_per_second) from the final
SSE data frame before yielding to the client. Streaming requests
get real throughput data instead of 0 tok/s.
Uses llama.cpp timings field in the last content chunk:
- completion_tokens = predicted_n
- tokens_per_sec = predicted_per_second
- inference_ms = predicted_ms (generation only)
Client sees identical stream, no perceptible delay.
- Dashboard: when a model has zero non-streaming records, shows
"streaming only" instead of misleading 0 tok/s
- Dashboard: minimum bar width enforced (6% avg, 4% p50) so
low-tps models are always visible
- Router: removed inflated streaming tps estimate (prompt tokens
skewed results for long conversations)
Fixes Dense model appearing to "register nothing" when Mumuni
sends mostly streaming requests.
router/router.py (+158 lines):
- store_perf_record(): captures queue_ms, inference_ms, prompt_tokens,
completion_tokens, tokens_per_sec per request in Redis
- Per-model, per-reason, per-agent rolling windows (last 200-500)
- /metrics/performance?window=N endpoint with percentiles (p50/p95/p99)
for latency, throughput, and queue time per model/reason/agent
- Queue time now surfaced in routing metadata and routes:recent
- Streaming requests tracked with estimated prompt tokens
nginx/nginx.conf:
- Added /metrics/ proxy pass to router_api
Enables model performance comparison and routing tier validation.
VRAM percentage no longer marks GPU as saturated.
Saturation is about slot availability (handled by is_gpu_busy()),
not memory usage. Added vram_warning boolean flag (≥95% threshold)
for informational monitoring without affecting routing decisions.
27B Dense now correctly shows healthy at 91% VRAM.
router/router.py:
- check_gpu_health() now accepts configurable timeouts (sidecar_timeout, gpu_timeout)
- /health and /v1/models endpoints use fast 1.5s/1s timeouts (non-blocking)
- /v1/models now calls check_gpu_health once per model instead of twice
- GPU_CONTEXT updated to 262144 across all models (turboquant upgrade)
- 27B max_concurrent reduced 2→1 (24GB VRAM saturated at 256K context)
docker-compose.yml:
- Router healthcheck timeout 5s→15s, interval 15s→30s
- Nginx healthcheck timeout 5s→15s, interval 15s→30s
Fixes dashboard hang when any GPU is unreachable.
- Saturated flag now triggers on load saturation (was dead code)
- Heavy tier routes MoE(131K) first instead of Dense(98K)
- Token estimation uses JSON length/3.5 (was content/4)
- Cross-turn session tracking via X-Session-Id + Redis TTL 24h
- Added GPU_CONTEXT map (MoE 131K, VLM 131K, Dense 65K)
- Heavy tier now prefers MoE/VLM (131K) over Dense (65K) for large requests
- Response headers: X-Context-Remaining, X-Context-Model
- Routing data includes context_remaining field
- Agents can use this to trigger compaction when nearing limits
Agent conversations with system prompts easily exceed 4000 tokens,
forcing everything to Dense. Now only truly heavy work triggers Dense.
Most agent convos will route to MoE (default) instead.
Previously it picked the least-loaded GPU globally, ignoring priority order.
Now it tries candidates in order: MoE → VLM → Dense. Only falls back to
least-loaded when ALL candidates are busy.
System messages are common in agent conversations but don't indicate
heavy workload. Now only token count (>4000) and turn count (>8) trigger
heavy routing. Simple conversations with system prompts can now route to VLM.
Mumuni's Ollama client probes /v1/props for model discovery and
/v1/models/<id> for per-model details. Previously both returned 404,
causing client retries. Now returns proper model properties and details.
When all GPUs are saturated, requests now enter a queue loop (poll every 500ms)
instead of immediately returning 503. Configurable via QUEUE_TIMEOUT env var
(default 30s) or X-Queue-Timeout header per-request.
This prevents agent failures from cluster saturation — agents wait for a slot
instead of crashing on fallback.
Abiba