feat(perf): verify shared recipe bundles and document pinned CLI workflow

README.md

@@ -78,6 +78,16 @@ target/release/grill-perf compare results/setup-a results/setup-b --json
 
 [Full speed benchmark guide](docs/performance/README.md)
 
+For separately qualified DeepSeek and GLM studies, use the
+[shared recipe workflow](docs/performance/SHARED-RECIPES.md): externally pinned
+source installation, offline `bundle verify`, an explicit precollection policy,
+and independent baseline/candidate/repeat acquisitions. `decide` evaluates the
+captured observed-envelope policy; successful `compare` output is not PASS.
+The shared workflow does not establish live qualification or cross-recipe
+equivalence. Publish only a
+[manually reviewed report](docs/performance/SHARED-REPORT-TEMPLATE.md), not private
+run evidence.
+
 ## Quality evaluation (WIP)
 
 Use **`grill`** to collect model answers and check them against task rules.

crates/grill-perf/examples/glm-decode-v1.json

@@ -0,0 +1,55 @@
+{
+  "version": 1,
+  "name": "glm-decode-v1",
+  "request": {
+    "profile": "vllm-fixed-v1",
+    "stream": true,
+    "output": { "tokens": 400, "mode": "exact" },
+    "cache": "observe",
+    "temperature_milli": 0,
+    "top_p_milli": 1000,
+    "seed": null,
+    "thinking_control": { "kind": "vllm-enable-thinking-v1", "enabled": false }
+  },
+  "limits": {
+    "total_ms": 360000,
+    "idle_ms": 60000,
+    "response_bytes": 1048576,
+    "wave_buffer_bytes": 67108864
+  },
+  "cases": [
+    {
+      "id": "structured",
+      "messages": [
+        { "role": "user", "content": "Count from 1 to 200. Output only the numbers, separated by spaces. No other text." }
+      ]
+    },
+    {
+      "id": "prose",
+      "messages": [
+        { "role": "user", "content": "Write a detailed step-by-step explanation of how a hash map works, including collision handling, resizing, and time complexity. Be thorough." }
+      ]
+    },
+    {
+      "id": "code",
+      "messages": [
+        { "role": "user", "content": "Output only Python source code. No comments, no docstrings, no markdown fences. Write functions clamp_00 through clamp_49. Each function is exactly:\ndef clamp_NN(x, lo=0, hi=1):\n    if x < lo:\n        return lo\n    if x > hi:\n        return hi\n    return x\nChange only the function name suffix (00, 01, … 49). One blank line between functions. No other text." }
+      ]
+    },
+    {
+      "id": "json",
+      "messages": [
+        { "role": "user", "content": "Emit only a JSON array of fake GPU metrics rows. Each object needs host, gpuIndex, utilPct, tempC, powerW, memUsedMb. Invent many rows. No markdown. Keep expanding the array." }
+      ]
+    }
+  ],
+  "cells": [
+    { "id": "structured-1", "case": "structured", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "structured-2", "case": "structured", "concurrency": 2, "warmup_trials": 1, "trials": 3 },
+    { "id": "structured-4", "case": "structured", "concurrency": 4, "warmup_trials": 1, "trials": 3 },
+    { "id": "structured-8", "case": "structured", "concurrency": 8, "warmup_trials": 1, "trials": 3 },
+    { "id": "prose-1", "case": "prose", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "code-1", "case": "code", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "json-1", "case": "json", "concurrency": 1, "warmup_trials": 1, "trials": 3 }
+  ]
+}

crates/grill-perf/examples/glm-prefill-v1.json

@@ -0,0 +1,56 @@
+{
+  "version": 1,
+  "name": "glm-prefill-v1",
+  "request": {
+    "profile": "vllm-fixed-v1",
+    "stream": true,
+    "output": { "tokens": 8, "mode": "exact" },
+    "cache": "observe",
+    "temperature_milli": 0,
+    "top_p_milli": 1000,
+    "seed": null,
+    "thinking_control": { "kind": "vllm-enable-thinking-v1", "enabled": false }
+  },
+  "limits": {
+    "total_ms": 600000,
+    "idle_ms": 600000,
+    "response_bytes": 65536,
+    "wave_buffer_bytes": 4194304
+  },
+  "cases": [
+    {
+      "id": "prefill-4k",
+      "messages": [
+        { "role": "user", "content": "[prefill-bench {salt}]\nIgnore the filler below. Reply with the single word OK.\n{fill}\nReply OK." }
+      ],
+      "fill": { "unit": " the", "repeat": 4070 }
+    },
+    {
+      "id": "prefill-8k",
+      "messages": [
+        { "role": "user", "content": "[prefill-bench {salt}]\nIgnore the filler below. Reply with the single word OK.\n{fill}\nReply OK." }
+      ],
+      "fill": { "unit": " the", "repeat": 8166 }
+    },
+    {
+      "id": "prefill-16k",
+      "messages": [
+        { "role": "user", "content": "[prefill-bench {salt}]\nIgnore the filler below. Reply with the single word OK.\n{fill}\nReply OK." }
+      ],
+      "fill": { "unit": " the", "repeat": 16358 }
+    },
+    {
+      "id": "prefill-32k",
+      "messages": [
+        { "role": "user", "content": "[prefill-bench {salt}]\nIgnore the filler below. Reply with the single word OK.\n{fill}\nReply OK." }
+      ],
+      "fill": { "unit": " the", "repeat": 32742 }
+    }
+  ],
+  "cells": [
+    { "id": "prefill-4k-1", "case": "prefill-4k", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "prefill-8k-1", "case": "prefill-8k", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "prefill-16k-1", "case": "prefill-16k", "concurrency": 1, "warmup_trials": 1, "trials": 3 },
+    { "id": "prefill-32k-1", "case": "prefill-32k", "concurrency": 1, "warmup_trials": 1, "trials": 3 }
+  ]
+}

crates/grill-perf/examples/recipes-v1.json

@@ -0,0 +1,41 @@
+{
+  "version": 1,
+  "entries": [
+    {
+      "id": "sparkdash-decode-v1",
+      "file": "sparkdash-decode-v1.json",
+      "source_sha256": "d661f6a6ff4718b8999b9796844345e9023b55b971b51d6fea1307e177eecdee",
+      "workload_sha256": "89dd5320c4f000cc7a15b1d352c721f1e26ddaf9911cccafe7818761d3acffef"
+    },
+    {
+      "id": "sparkdash-prefill-v1",
+      "file": "sparkdash-prefill-v1.json",
+      "source_sha256": "edf255fb30c20e934060ff08af1cfc2c680e6d1b4a2bf9417ce5d96389380468",
+      "workload_sha256": "549cf323c7117fd31097dc80ffdc267e6b748eab98f13d0a80b69f00b7274f3c"
+    },
+    {
+      "id": "glm-decode-v1",
+      "file": "glm-decode-v1.json",
+      "source_sha256": "508791d3e04a7cb7bc82f06084a5cac718834387123c826c7f378f43f912cb98",
+      "workload_sha256": "2129bcb5239d58d900a112ca5c2d511136292d54ab9d0c21dce6eaf9e37155a9",
+      "base": "sparkdash-decode-v1"
+    },
+    {
+      "id": "glm-prefill-v1",
+      "file": "glm-prefill-v1.json",
+      "source_sha256": "ffe1956899f4fc704f047fcaeb85733ebd100beff71676ca59b9bb7235ac16a7",
+      "workload_sha256": "31312a9a8ad58b7fb6d66bc927d79ce789a5bb6f31737ceb36d8e8d8d8d391f5",
+      "base": "sparkdash-prefill-v1"
+    }
+  ],
+  "recipes": {
+    "deepseek": {
+      "decode": "sparkdash-decode-v1",
+      "prefill": "sparkdash-prefill-v1"
+    },
+    "glm": {
+      "decode": "glm-decode-v1",
+      "prefill": "glm-prefill-v1"
+    }
+  }
+}