Files
Mortdecai df5542f7d6 feat: native-bakeoff scaffold — Ollama JSON vs native-token tool-calling
Three-arm harness under scripts/native-bakeoff/:
- arm A: /api/chat with JSON tools (current default)
- arm B: /api/generate raw:true with canonical HF jinja template rendered directly
- arm C: google-deepmind/gemma JAX ToolSampler (env-gated, JAX required)

Interim finding from A+B sweep on matt-strix gemma4:26b Q4: Ollama's
bidirectional JSON↔native tool-call translator is faithful. The "long"
multi-tool task produces identical behavior (7 steps / 6 tools) on both
arms. Earlier arm-B parser bug that looked like a divergence was a
harness issue: preserving the model's <|channel>thought\n<channel|>
prefix as assistant content tripped the jinja template's
tool_response-following conditional, appending a spurious <turn|>\n
that corrupted the next step's prompt. Fixed by dropping the channel
prefix on the assistant message.

Arm C left as scaffolded-but-not-run — the JAX/bf16 reference path
would answer "does the GGUF runtime diverge from DeepMind's
implementation" but requires a separate env with the `gemma` PyPI
package. Parked pending SDXL eviction or vast-h100 session.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-20 05:45:12 -04:00

276 lines
9.8 KiB
Python

"""Arm B: Ollama /api/generate with raw:true and native Gemma 4 tokens.
Renders the canonical HF jinja chat template directly, sends the
resulting string to Ollama's /api/generate with `raw: true` (which
bypasses Ollama's own templating / BOS handling), and parses
<|tool_call>call:NAME{args}<tool_call|> out of the completion with a
regex.
The point of this arm: isolate Ollama's tool parser. Arm A lets
Ollama's server translate OpenAI-shaped JSON tools into native tokens
AND translate the model's native <|tool_call> output back into
structured `tool_calls`. Arm B keeps everything native end-to-end and
only uses Ollama as a thin completion engine. If A and B diverge, the
delta lives in Ollama's bidirectional JSON↔native translator.
Template source: tooling/huggingface/model-cards/gemma-4-E4B-it-chat_template.jinja
"""
from __future__ import annotations
import asyncio
import json
import re
import time
from pathlib import Path
from typing import Any
import aiohttp
import jinja2
from tasks import SYSTEM_PROMPT, TOOLS, FAKE_HISTORY, execute_tool_stub
_REPO_ROOT = Path(__file__).resolve().parents[3]
_TEMPLATE_PATH = _REPO_ROOT / "tooling" / "huggingface" / "model-cards" / "gemma-4-E4B-it-chat_template.jinja"
def _load_template() -> jinja2.Template:
env = jinja2.Environment(
keep_trailing_newline=True,
# Canonical template uses `{%- ... -%}` whitespace control; keep
# jinja defaults so it renders exactly as HF's template expects.
)
return env.from_string(_TEMPLATE_PATH.read_text())
_TOOL_CALL_RE = re.compile(
r"<\|tool_call>call:(?P<name>\w+)\{(?P<body>.*?)\}<tool_call\|>",
re.DOTALL,
)
def _parse_native_args(body: str) -> dict[str, Any]:
"""Parse the body of a <|tool_call>call:NAME{...}<tool_call|>.
Gemma 4 native arg format (from the jinja template's format_argument
macro with escape_keys=False):
- key:<|"|>stringval<|"|>
- key:123
- key:true / key:false
- key:{nested:...} (for mapping args — not used by our stubs)
- key:[<|"|>item<|"|>,...] (for array args — not used by our stubs)
Our stub tool schemas are flat (string / integer / bool), so a
simple top-level comma split is enough. If a future tool needs
nested args this needs depth-aware splitting.
"""
out: dict[str, Any] = {}
if not body:
return out
# Top-level comma split, respecting only the `<|"|>...<|"|>` string
# delimiter (since our tool args don't nest). This intentionally
# doesn't handle {...} or [...] — flag it with a log entry in the
# harness if a future tool needs those.
parts: list[str] = []
buf = ""
i = 0
str_delim = '<|"|>'
in_str = False
while i < len(body):
if body[i : i + len(str_delim)] == str_delim:
in_str = not in_str
buf += str_delim
i += len(str_delim)
continue
if body[i] == "," and not in_str:
parts.append(buf)
buf = ""
i += 1
continue
buf += body[i]
i += 1
if buf:
parts.append(buf)
for p in parts:
if ":" not in p:
continue
k, _, v = p.partition(":")
k = k.strip()
v = v.strip()
if v.startswith(str_delim) and v.endswith(str_delim):
out[k] = v[len(str_delim) : -len(str_delim)]
elif v == "true":
out[k] = True
elif v == "false":
out[k] = False
else:
try:
out[k] = int(v)
except ValueError:
try:
out[k] = float(v)
except ValueError:
out[k] = v
return out
def _render(messages: list[dict[str, Any]]) -> str:
tmpl = _load_template()
return tmpl.render(
messages=messages,
tools=TOOLS,
add_generation_prompt=True,
bos_token="<bos>",
enable_thinking=False,
)
async def run(
*,
ollama_url: str,
model: str,
task_prompt: str,
num_ctx: int,
num_predict: int,
step_budget: int,
) -> dict[str, Any]:
messages: list[dict[str, Any]] = [{"role": "system", "content": SYSTEM_PROMPT}] + list(FAKE_HISTORY)
messages.append({"role": "user", "content": f"[2026-04-18 14:20] @seth:sethpc.xyz: {task_prompt}"})
trace: dict[str, Any] = {
"arm": "ollama-native",
"model": model,
"num_ctx": num_ctx,
"num_predict": num_predict,
"started_at": time.time(),
"turns": [],
"final": None,
}
tool_call_total = 0
halt: str | None = None
async with aiohttp.ClientSession() as session:
for step in range(1, step_budget + 1):
t0 = time.time()
prompt = _render(messages)
payload = {
"model": model,
"prompt": prompt,
"raw": True,
"stream": False,
"options": {
"num_ctx": num_ctx,
"num_predict": num_predict,
"temperature": 0.7, "top_p": 0.95, "top_k": 64,
# Stop at either end-of-turn (final answer) or end-of-tool-call.
# "<tool_call|>" lets the regex match on the full call; we
# re-append "<tool_call|>" before parsing to keep the regex
# simple. "<turn|>" catches a clean final answer.
"stop": ["<turn|>", "<tool_call|>"],
},
"keep_alive": "2h",
}
try:
async with session.post(
f"{ollama_url}/api/generate", json=payload,
timeout=aiohttp.ClientTimeout(total=300),
) as resp:
r = await resp.json()
except Exception as e:
halt = f"error: {e}"
trace["turns"].append({"step": step, "error": str(e)})
break
completion = r.get("response", "") or ""
stop_reason_native = r.get("done_reason") or r.get("stop_reason") or ""
# Rebuild the full assistant turn. Ollama's /api/generate
# strips the matched stop token from the response, so we
# always re-append based on which open token is present.
# An unclosed `<|tool_call>` means the model was emitting a
# tool call when the stop token fired; otherwise the model
# was producing a final text turn.
if completion.rstrip().endswith(("<tool_call|>", "<turn|>")):
full = completion
elif "<|tool_call>" in completion and "<tool_call|>" not in completion:
full = completion + "<tool_call|>"
else:
full = completion + "<turn|>"
matches = list(_TOOL_CALL_RE.finditer(full))
history_chars = sum(len(m.get("content", "") or "") for m in messages)
trace["turns"].append({
"step": step,
"elapsed_s": round(time.time() - t0, 2),
"prompt_eval_count": r.get("prompt_eval_count"),
"eval_count": r.get("eval_count"),
"content_len": len(completion),
"tool_call_count": len(matches),
"stop_reason": stop_reason_native,
"history_chars_before_append": history_chars,
"raw_completion_head": completion[:240],
"raw_completion_tail": completion[-240:] if len(completion) > 240 else "",
"prompt_tail": prompt[-400:],
"prompt_head": prompt[:200],
})
if not matches:
# Final answer — take the text minus any trailing <turn|>.
content = full.replace("<turn|>", "").strip()
messages.append({"role": "assistant", "content": content})
halt = "no_tool_calls"
break
# Build an assistant message with tool_calls (OpenAI shape) so the
# jinja template re-renders them correctly on the next iteration.
tool_calls_msg: list[dict[str, Any]] = []
for m in matches:
name = m.group("name")
args = _parse_native_args(m.group("body"))
tool_calls_msg.append({
"id": f"call_{step}_{len(tool_calls_msg)}",
"function": {"name": name, "arguments": args},
})
# Content MUST be empty when the message has tool_calls + will
# have tool_responses inlined on next render. The jinja
# template's post-turn conditional checks message.get('content')
# before strip_thinking and any non-empty string (even a bare
# <|channel>thought\n<channel|> prefix from the model) causes
# a spurious <turn|>\n to be appended after <tool_response|>,
# which breaks turn continuation on the following step.
messages.append({
"role": "assistant",
"content": "",
"tool_calls": tool_calls_msg,
})
tool_call_total += len(tool_calls_msg)
for tc in tool_calls_msg:
fn = tc["function"]
result = execute_tool_stub(fn["name"], fn["arguments"])
messages.append({
"role": "tool",
"tool_call_id": tc["id"],
"name": fn["name"],
"content": result,
})
if step == step_budget:
halt = "step_budget"
break
trace["final"] = {
"halt_reason": halt,
"steps_used": len(trace["turns"]),
"tool_calls_total": tool_call_total,
"wall_clock_s": round(time.time() - trace["started_at"], 2),
"final_message_count": len(messages),
"final_history_chars": sum(len(m.get("content", "") or "") for m in messages),
}
return trace