Files
esh-pfi-infrastructure/services/semif-serve/src/semif_serve/app.py
T
vh 069725c4b3 feat(semif): SemIf option-logit decisions on fv-ml1 GPU 1 (Prime)
services/semif-serve is a FastAPI wrapper around SemIf's direct and shared torch
scorers (SemIf-OpenJev @ 23cf1f39, MIT). Upstream ships only a batch CLI. The
wrapper loads the pinned Qwen3.5-4B (851bf6e8, BF16) once from the offline HF
cache and returns SemIf's result dicts unchanged, with an optional per-workload
temperature-calibrated view. Contract: semif-serve.contract.md. Built with a
short contract, TDD (39 tests, fake engine and fake torch, no GPU) and a heid
bug-hunt panel (pending).

On the card:
- torch 2.10.0+cu128 with sm_120 kernels, which is SemIf's own stack;
- a hard 12 GiB VRAM cap.
Two defects surfaced only on the card, and each fix is covered by a test:
- 0.1.1: an OOM raised as a chained exception kept the failed request's tensors
  alive (11.9 GiB after the 503). It is now raised unchained, after gc.
- 0.1.2: a large request left 12.6 GB reserved on the shared card. After each
  call, reserved memory over the baseline + 512 MiB is now released.

Acceptance against SemIf's committed torch predictions (authored144):
- 142/144 same top choice; both misses are exact bf16 ties;
- 144/144 identical prompt hashes;
- deterministic A-vs-A;
- negative control 14/144;
- shared vs direct 72/72.
21 binary criteria over one state take 159 ms. The shared-mode capacity table
under the cap is in stacks/semif/README.md.

The Dockerfile installs dependencies from a manifest with the project version
blanked, so a version bump reuses the ~4 GB torch layer. Verified: 41 s rebuild,
dependency layer CACHED.

DNS: semif.fv.internal. Token: vault semif/api-token.
2026-09-27 02:36:56 -07:00

162 lines
6.3 KiB
Python

"""semif-serve HTTP layer. Contract: semif-serve.contract.md."""
from __future__ import annotations
import hmac
import math
import threading
from typing import Any
from fastapi import FastAPI, Request
from fastapi.concurrency import run_in_threadpool
from fastapi.responses import JSONResponse
from pydantic import BaseModel, ConfigDict, ValidationError
from .config import SEMIF_COMMIT, Settings
from .errors import OutOfMemory
OPEN_PATHS = frozenset({"/health"})
State = str | dict[str, Any] | list[Any]
class Option(BaseModel):
model_config = ConfigDict(extra="ignore")
id: str
description: str
class Decision(BaseModel):
model_config = ConfigDict(extra="ignore")
id: str
question: str
options: list[Option]
def row(self, state: State) -> dict:
"""The SemIf row shape: exactly id, state, question, options."""
return {"id": self.id, "state": state, "question": self.question,
"options": [o.model_dump() for o in self.options]}
class DecideBody(Decision):
state: State
workload: str | None = None
class SharedBody(BaseModel):
model_config = ConfigDict(extra="ignore")
state: State
decisions: list[Decision]
workload: str | None = None
class ApiError(Exception):
def __init__(self, status: int, code: str, message: str):
super().__init__(message)
self.status, self.code, self.message = status, code, message
def error(status: int, code: str, message: str) -> JSONResponse:
return JSONResponse(status_code=status, content={"error": {"code": code, "message": message}})
def _first_error(exc: ValidationError) -> str:
first = exc.errors()[0]
where = ".".join(str(p) for p in first.get("loc", ())) or "body"
return f"{where}: {first.get('msg', 'invalid')}"
def calibrated_view(result: dict, workload: str, temperature: float) -> dict:
"""softmax(option_logits / T): the native fields are left exactly as SemIf returned them (INV-1)."""
scaled = [x / temperature for x in result["option_logits"]]
top = max(scaled)
weights = [math.exp(x - top) for x in scaled]
total = sum(weights)
return {"workload": workload, "temperature": temperature, "probabilities": [w / total for w in weights]}
def create_app(settings: Settings, engine: Any) -> FastAPI:
app = FastAPI(title="semif-serve")
expected = f"Bearer {settings.api_token}".encode()
inference = threading.Lock() # INV-2: one scorer call at a time, off the event loop
def locked(fn, *args):
with inference:
return fn(*args)
@app.middleware("http")
async def require_bearer(request: Request, call_next):
if request.url.path not in OPEN_PATHS:
supplied = request.headers.get("authorization", "").encode()
if not hmac.compare_digest(supplied, expected): # INV-6
return error(401, "unauthorized", "missing or wrong bearer token")
return await call_next(request)
@app.exception_handler(ApiError)
async def api_error(_request: Request, exc: ApiError):
return error(exc.status, exc.code, exc.message)
async def read_limited(request: Request) -> bytes:
limit = settings.max_body_bytes
too_large = ApiError(413, "request_too_large", f"request body exceeds {limit} bytes")
declared = request.headers.get("content-length")
if declared is not None and declared.isdigit() and int(declared) > limit:
raise too_large
body = bytearray()
async for chunk in request.stream(): # also caps bodies that declare no length
body.extend(chunk)
if len(body) > limit:
raise too_large
return bytes(body)
async def parse(request: Request, model: type[BaseModel]):
try:
return model.model_validate_json(await read_limited(request))
except ValidationError as exc:
raise ApiError(422, "invalid_request", _first_error(exc)) from exc
async def score(fn, *args):
"""Run one scorer call in a worker thread under the lock; map its failures to contract codes."""
try:
return await run_in_threadpool(locked, fn, *args)
except ValueError as exc: # SemIf validation, token limit, tokenisation
raise ApiError(422, "invalid_request", str(exc)) from exc
except OutOfMemory as exc:
raise ApiError(503, "out_of_memory", str(exc)) from exc
except Exception as exc: # noqa: BLE001 — any other scorer failure
raise ApiError(500, "scoring_failed", f"{type(exc).__name__}: {exc}") from exc
def temperature_for(workload: str | None) -> float | None:
if workload is None:
return None
if workload not in settings.calibration:
raise ApiError(422, "invalid_request", f"unknown workload {workload!r}")
return settings.calibration[workload]
def with_calibration(result: dict, workload: str | None, temperature: float | None) -> dict:
if temperature is None:
return result
return {**result, "calibrated": calibrated_view(result, workload, temperature)}
@app.get("/health")
async def health():
return {"status": "ok", "semif_commit": SEMIF_COMMIT, "model": engine.health(),
"vram_cap_gib": settings.vram_cap_gib, "max_tokens": settings.max_tokens,
"max_decisions": settings.max_decisions, "workloads": sorted(settings.calibration)}
@app.post("/decide")
async def decide(request: Request):
body = await parse(request, DecideBody)
temperature = temperature_for(body.workload)
return with_calibration(await score(engine.direct, body.row(body.state)), body.workload, temperature)
@app.post("/decide/shared")
async def decide_shared(request: Request):
body = await parse(request, SharedBody)
if not 1 <= len(body.decisions) <= settings.max_decisions:
raise ApiError(422, "invalid_request",
f"decisions must hold 1..{settings.max_decisions} entries, got {len(body.decisions)}")
temperature = temperature_for(body.workload)
results, timing = await score(engine.shared, [d.row(body.state) for d in body.decisions])
return {"results": [with_calibration(r, body.workload, temperature) for r in results], "timing": timing}
return app