feat: CPU reasoner bake-off for Qwen3.5-9B vs Bonsai (#25)
Tests / Test (push) Failing after 5s
Tests / Release (semver) (push) Skipped

Measure OpenAI tool_calls (search/get/audit) on a CPU Ollama sidecar
instead of gating D18 on GPU or PicoClaw. Weights stay out of the image.
This commit is contained in:
2026-08-13 23:19:51 +01:00
committed by GitHub
co-authored by GitHub
parent e393cc6a99
commit bb272d7416
13 changed files with 717 additions and 3 deletions
+91
View File
@@ -0,0 +1,91 @@
//usr/bin/env go run -tags=reasoner_bakeoff "$0" "$@"; exit
//go:build reasoner_bakeoff
//
// bin/reasoner/bakeoff.go - CPU tool-call bake-off against an OpenAI-compatible URL (D18).
//
// REASONER_BASE_URL=http://127.0.0.1:11435/v1 REASONER_MODEL=qwen3.5:9b ./bin/reasoner/bakeoff.go
// ./bin/reasoner/bakeoff.go --model MichelRosselli/bonsai-27b:Q1_0 --json
//
// Measures OpenAI tool_calls (search/get/audit) and RSS from Ollama /api/ps, not VRAM.
// PicoClaw is not in this repo; the tool names match internal/httpapi MCP ops.
// NOTE: never run `gofmt -w` on this file — it breaks the shebang.
package main
import (
"encoding/json"
"fmt"
"os"
"strings"
"github.com/eSlider/2dph/internal/reasoner"
)
func main() {
os.Exit(run(os.Args[1:]))
}
func run(args []string) int {
base := os.Getenv("REASONER_BASE_URL")
if base == "" {
base = "http://127.0.0.1:11435/v1"
}
model := os.Getenv("REASONER_MODEL")
if model == "" {
model = reasoner.OllamaRAM
}
jsonOut := false
device := "cpu"
for i := 0; i < len(args); i++ {
a := args[i]
switch {
case a == "--json":
jsonOut = true
case a == "--model" && i+1 < len(args):
i++
model = args[i]
case strings.HasPrefix(a, "--model="):
model = strings.TrimPrefix(a, "--model=")
case a == "--base-url" && i+1 < len(args):
i++
base = args[i]
case a == "--device" && i+1 < len(args):
i++
device = args[i]
case a == "-h" || a == "--help":
fmt.Fprintln(os.Stderr, "bin/reasoner/bakeoff.go [--model ID] [--base-url URL] [--device cpu] [--json]")
return 0
default:
fmt.Fprintln(os.Stderr, "unknown arg:", a)
return 2
}
}
c := reasoner.Client{BaseURL: base, Model: model, Device: device}
rep := reasoner.Run(c)
raw, err := json.MarshalIndent(rep, "", " ")
if err != nil {
fmt.Fprintln(os.Stderr, err)
return 1
}
if jsonOut {
fmt.Println(string(raw))
} else {
fmt.Printf("model: %s\n", rep.Model)
fmt.Printf("hf_id: %s\n", rep.HF)
fmt.Printf("device: %s\n", rep.Device)
fmt.Printf("tool_call: %d/%d\n", rep.ToolCallOK, rep.ToolCallN)
fmt.Printf("xml_leak: %d\n", rep.XMLLeak)
fmt.Printf("rss_mb: %d\n", rep.RSSMB)
fmt.Printf("vram_mb: %d\n", rep.VRAMMB)
for _, p := range rep.Prompts {
status := "fail"
if p.OK {
status = "ok"
}
fmt.Printf(" %s: %s wanted=%s got=%s xml=%v %dms %s\n", p.WantedTool, status, p.WantedTool, p.ToolName, p.XMLLeak, p.LatencyMS, p.Err)
}
}
if rep.ToolCallN == 0 {
return 1
}
return 0
}
+2
View File
@@ -0,0 +1,2 @@
// Commands in this directory are shebang mains (bakeoff.go).
package main
+29
View File
@@ -103,6 +103,35 @@ class PublishedDocsTest(unittest.TestCase):
self.assertIn('profiles: ["index"]', compose)
self.assertIn("target: api", compose)
def test_reasoner_docs_name_real_hf_ids_cpu_sidecar(self) -> None:
docs = (ROOT / "docs" / "reasoner.md").read_text()
for hf in (
"Qwen/Qwen3.5-9B",
"Qwen/Qwen3.6-27B",
"prism-ml/Bonsai-27B-gguf",
):
self.assertIn(hf, docs)
self.assertIn("no official qwen3.6-9b", docs.lower())
self.assertIn("OLLAMA_NUM_GPU", docs)
self.assertIn("rss_mb", docs)
self.assertIn("vram_mb", docs)
self.assertIn("3/3", docs)
self.assertIn("Do not claim 9B is better at tools", docs)
self.assertNotIn("Qwen/Qwen3.6-9B", docs)
plan = (ROOT / "PLAN.md").read_text()
self.assertIn("D18", plan)
self.assertIn("Qwen/Qwen3.5-9B", plan)
compose = (ROOT / "compose.yaml").read_text()
self.assertIn('profiles: ["reasoner"]', compose)
self.assertIn("OLLAMA_NUM_GPU", compose)
self.assertIn("127.0.0.1:11435", compose)
dockerfile = (ROOT / "Dockerfile").read_text()
self.assertNotIn(".gguf", dockerfile.lower())
self.assertNotIn(".safetensors", dockerfile.lower())
api = dockerfile[dockerfile.index("FROM debian:bookworm-slim AS api") :]
self.assertNotIn("COPY models", api)
self.assertNotIn("qwen", api.lower())
def test_readme_search_escalates_web(self) -> None:
text = (ROOT / "README.md").read_text()
self.assertIn("--no-web", text)