File size: 12,286 Bytes
0366d65 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 | """Load the fine-tuned Gemma 4 GGUF and run inference via llama.cpp.
Llama Champion: all generation goes through llama-cpp-python — no cloud AI API.
The GGUF is downloaded from HF at startup so the Space image stays small.
Two inference locations, selected by env:
- in-process llama.cpp, GPU-offloaded inside an @spaces.GPU lease (ZeroGPU), or
- a remote OpenAI-compatible / llama.cpp server via INFERENCE_BASE_URL
(e.g. a llama-server on the phone itself, or a backend).
"""
from __future__ import annotations
import os
import threading
import time
from huggingface_hub import hf_hub_download
from . import events
# The platform runs the gemma-cal EDGE fine-tune (Gemma-4 E4B, ~5GB Q4) — our own
# calendar-native model, eval-gated before every publish (docs/eval-roadmap.md).
# MODEL SIZE (hackathon hard constraint, <= 32B): E4B = ~4B effective params.
# All inference is local via llama.cpp (no cloud AI).
MODEL_REPO = os.environ.get("MODEL_REPO", "ParetoOptimal/gemma-4-cal-gguf")
MODEL_FILE = os.environ.get("MODEL_FILE", "gemma-cal-e4b-Q4_K_M.gguf")
# Vision projector (mmproj). Set to enable image input; leave empty for text-only.
# MMPROJ_REPO lets the projector come from a different repo than the LLM — the E4B
# edge model pairs with the base E4B's projector, not a projector in our repo.
MMPROJ_REPO = os.environ.get("MMPROJ_REPO", "") or os.environ.get("MODEL_REPO", "ParetoOptimal/gemma-4-cal-gguf")
MMPROJ_FILE = os.environ.get("MMPROJ_FILE", "")
# llama-cpp-python vision handler class (in llama_cpp.llama_chat_format). Gemma 4
# vision may ship a dedicated handler; the generic clip/Llava handler is the default.
CHAT_HANDLER = os.environ.get("CHAT_HANDLER", "Llava15ChatHandler")
N_CTX = int(os.environ.get("N_CTX", "8192"))
N_GPU_LAYERS = int(os.environ.get("N_GPU_LAYERS", "-1")) # -1 = offload all (GPU)
GPU_DURATION = int(os.environ.get("GPU_DURATION", "120")) # ZeroGPU lease seconds
# Configurable inference location. If INFERENCE_BASE_URL is set, generation is
# delegated to a remote OpenAI-compatible / llama.cpp server (e.g. a llama-server
# running on the phone itself, or a backend) instead of loading the GGUF in-process.
# This is how the same agent runs on-device OR thin-client — selected by env.
INFERENCE_BASE_URL = os.environ.get("INFERENCE_BASE_URL", "")
INFERENCE_API_KEY = os.environ.get("INFERENCE_API_KEY", "")
INFERENCE_MODEL = os.environ.get("INFERENCE_MODEL", "local")
# Let a tool-calling model (Hermes) write its own long-term memory mid-run.
# Only applies to the remote path (server/tools.py); off by default.
HERMES_TOOLS = os.environ.get("HERMES_TOOLS") == "1"
_llm = None
_lock = threading.Lock()
# ZeroGPU: GPU-bound work must run inside an @spaces.GPU function (the GPU is
# attached only for that call). Locally / in CI the `spaces` package is absent,
# so `gpu` degrades to a no-op decorator and stub mode never touches this path.
try:
from spaces import GPU as _spaces_gpu
def gpu(fn):
return _spaces_gpu(duration=GPU_DURATION)(fn)
except Exception: # noqa: BLE001 - spaces not installed (local/CI)
def gpu(fn):
return fn
def _preload_cuda_libs():
"""Preload CUDA userspace libs so the prebuilt CUDA llama-cpp-python wheel can
dlopen. The ZeroGPU/Gradio-SDK env lacks libcudart.so.12 on the default loader
path; the nvidia-*-cu12 pip packages provide them. We CDLL them RTLD_GLOBAL so
the llama .so's NEEDED deps resolve. Path-independent (no LD_LIBRARY_PATH guess);
a no-op off-Linux / when the packages aren't installed."""
import ctypes
import glob
import os
try:
import nvidia # namespace package from nvidia-*-cu12 wheels
except Exception: # noqa: BLE001
return
# nvidia is a PEP 420 namespace package: __file__ is None, use __path__.
bases = list(getattr(nvidia, "__path__", []) or [])
# cublas before its dependents is unnecessary ($ORIGIN RPATH resolves siblings).
for base in bases:
for sub in ("cuda_runtime", "cuda_nvrtc", "cublas"):
for so in sorted(glob.glob(os.path.join(base, sub, "lib", "*.so*"))):
try:
ctypes.CDLL(so, mode=ctypes.RTLD_GLOBAL)
except OSError:
pass
def _build_chat_handler():
"""Return a vision chat handler if MMPROJ_FILE is set, else None (text-only)."""
if not MMPROJ_FILE:
return None
import llama_cpp.llama_chat_format as fmt
mmproj_path = hf_hub_download(repo_id=MMPROJ_REPO, filename=MMPROJ_FILE)
handler_cls = getattr(fmt, CHAT_HANDLER)
return handler_cls(clip_model_path=mmproj_path, verbose=False)
def get_llm():
"""Lazily download + load the GGUF once, thread-safe."""
global _llm
if _llm is None:
with _lock:
if _llm is None:
_preload_cuda_libs() # satisfy libcudart.so.12 etc. before loading
from llama_cpp import Llama # imported lazily so tests can stub
path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
_llm = Llama(
model_path=path,
n_ctx=N_CTX,
n_gpu_layers=N_GPU_LAYERS,
chat_handler=_build_chat_handler(), # enables image_url inputs
verbose=False,
)
return _llm
# --- GPU-scoped inner functions (run inside the ZeroGPU lease) ---
# These do the actual in-process llama.cpp work; emits stay in the main-process
# wrappers below because in-memory state (the events bus) isn't shared back from
# the ZeroGPU subprocess.
@gpu
def _infer_text(messages: list[dict], temperature: float, max_tokens: int) -> str:
out = get_llm().create_chat_completion(
messages=messages, temperature=temperature, max_tokens=max_tokens
)
return out["choices"][0]["message"]["content"]
@gpu
def _infer_json(messages: list[dict], json_schema: dict, temperature: float, max_tokens: int):
out = get_llm().create_chat_completion(
messages=messages,
temperature=temperature,
max_tokens=max_tokens,
response_format={"type": "json_object", "schema": json_schema},
)
usage = out.get("usage") or {}
return out["choices"][0]["message"]["content"], usage.get("completion_tokens")
@gpu
def _infer_stream(messages: list[dict], json_schema: dict, temperature: float, max_tokens: int):
stream = get_llm().create_chat_completion(
messages=messages,
temperature=temperature,
max_tokens=max_tokens,
response_format={"type": "json_object", "schema": json_schema},
stream=True,
)
for chunk in stream:
delta = chunk["choices"][0].get("delta", {}).get("content")
if delta:
yield delta
# --- remote inference seam (on-device / thin-client via INFERENCE_BASE_URL) ---
def _remote_payload(messages, json_schema, temperature, max_tokens, stream):
return {
"model": INFERENCE_MODEL,
"messages": messages,
"temperature": temperature,
"max_tokens": max_tokens,
# llama-server accepts json_schema (OpenAI-style); the in-process path uses
# the json_object+schema form. Both grammar-constrain the output.
"response_format": {
"type": "json_schema",
"json_schema": {"name": "ActionPlan", "schema": json_schema, "strict": True},
},
"stream": stream,
}
def _remote_headers() -> dict:
h = {"Content-Type": "application/json"}
if INFERENCE_API_KEY:
h["Authorization"] = f"Bearer {INFERENCE_API_KEY}"
return h
def _remote_complete_json(messages, json_schema, temperature, max_tokens) -> str:
import requests # already a dependency; imported here to keep import light
t0 = time.perf_counter()
if HERMES_TOOLS:
# Tool-calling loop: the model may call `remember` to update memory before
# returning the final ActionPlan JSON. See server/tools.py.
from .tools import TOOL_SPECS, run_with_tools
def _post(msgs):
payload = _remote_payload(msgs, json_schema, temperature, max_tokens, False)
payload["tools"] = TOOL_SPECS
r = requests.post(
f"{INFERENCE_BASE_URL.rstrip('/')}/chat/completions",
json=payload,
headers=_remote_headers(),
timeout=120,
)
r.raise_for_status()
return r.json()
content, out = run_with_tools(list(messages), _post)
usage = out.get("usage") or {}
events.emit(
"model",
"remote inference complete (tools)",
latency_ms=round((time.perf_counter() - t0) * 1000),
tokens=usage.get("completion_tokens"),
)
return content
resp = requests.post(
f"{INFERENCE_BASE_URL.rstrip('/')}/chat/completions",
json=_remote_payload(messages, json_schema, temperature, max_tokens, False),
headers=_remote_headers(),
timeout=120,
)
resp.raise_for_status()
out = resp.json()
usage = out.get("usage") or {}
events.emit(
"model",
"remote inference complete",
latency_ms=round((time.perf_counter() - t0) * 1000),
tokens=usage.get("completion_tokens"),
)
return out["choices"][0]["message"]["content"]
def _remote_stream_json(messages, json_schema, temperature, max_tokens):
import json as _json
import requests
t0 = time.perf_counter()
events.emit("model", "remote inference started")
with requests.post(
f"{INFERENCE_BASE_URL.rstrip('/')}/chat/completions",
json=_remote_payload(messages, json_schema, temperature, max_tokens, True),
headers=_remote_headers(),
timeout=120,
stream=True,
) as resp:
resp.raise_for_status()
for raw in resp.iter_lines():
if not raw:
continue
line = raw.decode("utf-8").removeprefix("data: ").strip()
if not line or line == "[DONE]":
continue
try:
delta = _json.loads(line)["choices"][0].get("delta", {}).get("content")
except (ValueError, KeyError, IndexError):
continue
if delta:
yield delta
events.emit(
"model", "remote stream complete", latency_ms=round((time.perf_counter() - t0) * 1000)
)
# --- main-process wrappers (own the activity-bus emits; pick local vs remote) ---
def complete(messages: list[dict], temperature: float = 0.2, max_tokens: int = 1024) -> str:
"""Chat-completion helper returning the assistant text."""
return _infer_text(messages, temperature, max_tokens)
def complete_json(
messages: list[dict],
json_schema: dict,
temperature: float = 0.2,
max_tokens: int = 2048,
) -> str:
"""Constrained completion: grammar-constrained so the output always parses.
Delegates to a remote server if INFERENCE_BASE_URL is set, else runs the
GPU-offloaded in-process llama.cpp path."""
if INFERENCE_BASE_URL:
return _remote_complete_json(messages, json_schema, temperature, max_tokens)
t0 = time.perf_counter()
text, tokens = _infer_json(messages, json_schema, temperature, max_tokens)
events.emit(
"model",
"inference complete",
latency_ms=round((time.perf_counter() - t0) * 1000),
tokens=tokens,
)
return text
def stream_complete_json(
messages: list[dict],
json_schema: dict,
temperature: float = 0.2,
max_tokens: int = 2048,
):
"""Streaming constrained completion: yields text deltas so the UI can show the
model 'thinking'. Remote seam when INFERENCE_BASE_URL is set, else GPU-offloaded
in-process llama.cpp. Emits model events around the call."""
if INFERENCE_BASE_URL:
yield from _remote_stream_json(messages, json_schema, temperature, max_tokens)
return
t0 = time.perf_counter()
events.emit("model", "inference started")
for delta in _infer_stream(messages, json_schema, temperature, max_tokens):
yield delta
events.emit(
"model", "stream complete", latency_ms=round((time.perf_counter() - t0) * 1000)
)
|