@remnic/bench 9.3.681 → 9.3.683

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/bench",
-  "version": "9.3.681",
+  "version": "9.3.683",
   "description": "Retrieval latency ladder benchmarks + CI regression gates for @remnic/core",
   "type": "module",
   "main": "./dist/index.js",
@@ -12,6 +12,8 @@
     },
     "./baselines/*": "./baselines/*",
     "./baselines/*.json": "./baselines/*.json",
+    "./profiles/*": "./profiles/*",
+    "./profiles/*.json": "./profiles/*.json",
     "./package.json": "./package.json"
   },
   "license": "MIT",
@@ -34,7 +36,7 @@
   "dependencies": {
     "hyparquet": "^1.25.7",
     "yaml": "^2.4.2",
-    "@remnic/core": "^9.3.681"
+    "@remnic/core": "^9.3.683"
   },
   "devDependencies": {
     "tsup": "^8.5.1",
@@ -42,7 +44,8 @@
   },
   "files": [
     "dist",
-    "baselines"
+    "baselines",
+    "profiles"
   ],
   "scripts": {
     "prebuild": "node ../../scripts/ensure-bench-build-deps.mjs",

package/profiles/README.md

@@ -0,0 +1,113 @@
+# Local-lab runtime profiles
+
+This directory ships reference JSON manifests for the **`local-lab`** bench
+runtime profile (issue #1573 PR2). A local-lab profile pins responder,
+judge, and (optionally) embedding to operator-hosted models so the bench
+runs entirely on local hardware — no paid API calls, no rate-limit cliffs.
+
+These manifests are **templates**. They use placeholder model ids on
+purpose: the operator must replace them with the ids their endpoint
+actually reports (`GET /v1/models` for OpenAI-compatible servers,
+`GET /api/tags` for Ollama) before running. The bench refuses to start a
+phase whose preflight cannot find the manifest model id (rule 51 — no
+silent fallback).
+
+## Schema
+
+|Field|Type|Meaning|
+|---|---|---|
+|`profile`|`"local-lab"`|Discriminator. Required.|
+|`responder`|role|Answering model (must report this id at the endpoint).|
+|`judge`|role|Scoring model. Runs in the judge phase after the responder.|
+|`embedding`|role?|Optional embedding model for retrieval indexing.|
+|`phases`|`"sequential"`|Phase scheduling mode. PR2 ships `sequential` only.|
+|`notes.responderToJudgeHandoff`|string?|Operator guidance for multi-endpoint swaps. PR2 ships same-endpoint profiles only (multi-endpoint sequential execution is PR3 scope).|
+
+Each role has shape:
+
+|Field|Type|Meaning|
+|---|---|---|
+|`provider`|`"openai-compatible"` \| `"ollama"`|Transport the bench uses for preflight + completion.|
+|`baseUrl`|string|Endpoint URL. Fetch target only — never interpolated into a shell.|
+|`model`|string|Exact id the endpoint reports. No aliases.|
+|`quantization`|string?|Informational; recorded in artifacts.|
+|`ctx`|positive int|Manifest-declared serving context (tokens). Preflight verifies the live endpoint reports at least this much.|
+|`temperature`|`0`|Pinned to 0 for reproducibility. Non-zero is rejected.|
+|`seed`|int|Sampling seed; required so reruns reproduce the same draws.|
+
+## `local-lab-3090.json`
+
+Reference profile targeting a single-GPU lab box. The numbers below come
+from issue #1573's *Hardware envelope* section; they are VRAM math, not
+performance claims, and are reproduced here so the manifest's defaults
+can be checked against the design constraints.
+
+### Hardware envelope (RTX 3090, 24 GB VRAM, Ampere; 256 GB RAM, NVMe)
+
+|Class|Rough VRAM @ 4-bit|Practical serving ctx|Notes|
+|---|---|---|---|
+|~14B dense (e.g. Qwen 2.5 14B, Llama 3 8B-class step-up)|9–10 GB|32k KV cache headroom|Comfortably fits; responder-only or judge-only.|
+|~24–32B dense or ~30B-class MoE|14–19 GB|≤16k ctx (tight)|Responder OR judge, not both; KV budget tight.|
+|Embedding model (e.g. BGE-M3, Qwen 3 embedding)|1–3 GB|8k ctx|Fits alongside one of the above on the same GPU in some setups.|
+
+Constraints the profile respects:
+
+- **24 GB VRAM fits one serious model at a time** — responder and judge
+  MUST run in sequential phases, never co-resident.
+- **Ampere: bf16 OK, no FP8.** Prefer AWQ/GPTQ-int4 or GGUF Q4_K_M
+  quantizations.
+- Remnic's memory-based answering keeps prompts small (that is the
+  product); LongMemEval_S's ~115k-token histories are ingested through
+  observe/extraction, not stuffed in context — so **16–32k serving context
+  is sufficient**.
+- **256 GB RAM**: model files stay in page cache; phase swaps from NVMe
+  are seconds, not minutes. The operator can hot-swap models between
+  phases without long stalls.
+
+### Defaults chosen in this manifest
+
+The reference manifest pins both `responder` and `judge` to a **16 384
+token** serving context. That value:
+
+- Lies inside the 16–32k envelope above.
+- Is conservatively below the "tight" ceiling for 24–32B dense models.
+- Matches what GGUF Q4_K_M builds of mainstream ~14B / ~30B-class models
+  expose out of the box, so an operator swapping in their own model id is
+  unlikely to need to lower `ctx`.
+
+`temperature: 0` and a fixed `seed: 1573` make reruns reproducible (the
+manifest parser rejects non-zero temperatures and missing seeds).
+
+The shipped profile pins all roles to a single Ollama endpoint
+(`http://127.0.0.1:11434`). On a single 3090 the operator runs one
+`ollama serve` instance and hot-swaps models between phases — no manual
+server restart is needed. **Multi-endpoint manifests** (responder and
+judge on different URLs) are not yet supported by the benchmark runner:
+the published harness executes recall→answer→judge per trial, which
+requires both models to be co-resident. Full sequential phase execution
+(answer-all-then-judge-all with an operator swap between) is tracked for
+PR3's calibration work.
+
+## Using a profile
+
+Resolution goes through `resolveBenchRuntimeProfile` (see
+`packages/bench/src/runtime-profiles.ts`); the resolver turns each role
+into a `ProviderConfig` with `temperature` and `seed` forwarded verbatim.
+At run time, the CLI preflights both responder and judge endpoints before
+any trial starts. Neither layer manages model processes — that is the
+operator's responsibility. The sequential phase scheduler
+(`runSequentialPhases`) is implemented and tested but not yet wired into
+the benchmark trial loop (PR3 calibration scope).
+
+```ts
+import {
+  loadLocalLabManifest,
+  resolveLocalLabProfile,
+  runSequentialPhases,
+} from "@remnic/bench";
+
+const manifest = await loadLocalLabManifest("packages/bench/profiles/local-lab-3090.json");
+const profile = resolveLocalLabProfile(manifest);
+// profile.responder.providerConfig.temperature === 0
+// profile.responder.providerConfig.seed === 1573
+```

package/profiles/local-lab-3090.json

@@ -0,0 +1,39 @@
+{
+  "profile": "local-lab",
+  "responder": {
+    "provider": "ollama",
+    "baseUrl": "http://127.0.0.1:11434",
+    "model": "PLACEHOLDER-RESPONDER-MODEL",
+    "quantization": "Q4_K_M",
+    "ctx": 16384,
+    "temperature": 0,
+    "seed": 1573
+  },
+  "judge": {
+    "provider": "ollama",
+    "baseUrl": "http://127.0.0.1:11434",
+    "model": "PLACEHOLDER-JUDGE-MODEL",
+    "quantization": "Q4_K_M",
+    "ctx": 16384,
+    "temperature": 0,
+    "seed": 1573
+  },
+  "embedding": {
+    "provider": "ollama",
+    "baseUrl": "http://127.0.0.1:11434",
+    "model": "PLACEHOLDER-EMBEDDING-MODEL",
+    "ctx": 8192,
+    "temperature": 0,
+    "seed": 1573
+  },
+  "phases": "sequential",
+  "notes": {
+    "responderToJudgeHandoff": "Both responder and judge models run on the same Ollama instance (hot-swap). Ensure `ollama pull` has both PLACEHOLDER-RESPONDER-MODEL and PLACEHOLDER-JUDGE-MODEL before starting the bench.",
+    "hardware": {
+      "gpu": "NVIDIA RTX 3090",
+      "vramGb": 24,
+      "ramGb": 256,
+      "ampereBf16Ok": true
+    }
+  }
+}