GradLLM / hf_backend.py
johnbridges's picture
.
1175344
import time, logging, json, asyncio
from contextlib import nullcontext
from typing import Any, Dict, AsyncIterable, Tuple
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
from backends_base import ChatBackend, ImagesBackend
from config import settings
logger = logging.getLogger(__name__)
def _snippet(txt: str, n: int = 800) -> str:
if not isinstance(txt, str):
return f"<non-str:{type(txt)}>"
return txt if len(txt) <= n else txt[:n] + f"... <+{len(txt)-n} chars>"
try:
import spaces
from spaces.zero import client as zero_client
except ImportError:
spaces, zero_client = None, None
MODEL_ID = settings.LlmHFModelID or "Qwen/Qwen2.5-1.5B-Instruct"
logger.info(f"[init] MODEL_ID={MODEL_ID}")
tokenizer, load_error = None, None
try:
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, use_fast=False)
has_template = hasattr(tokenizer, "apply_chat_template") and getattr(tokenizer, "chat_template", None)
logger.info(f"[init] tokenizer loaded. chat_template={'yes' if has_template else 'no'}")
except Exception as e:
load_error = f"Failed to load tokenizer: {e}"
logger.exception(load_error)
def probe_bf16_runtime() -> bool:
"""Check if BF16 is both reported and actually used in ops on CPU."""
if not (hasattr(torch, "cpu") and hasattr(torch.cpu, "is_bf16_supported")):
return False
if not torch.cpu.is_bf16_supported():
return False
try:
a = torch.randn(16, 16, dtype=torch.bfloat16)
b = torch.randn(16, 16, dtype=torch.bfloat16)
c = a @ b
return c.dtype == torch.bfloat16
except Exception:
return False
def _pick_cpu_dtype() -> torch.dtype:
try:
if probe_bf16_runtime():
logger.info("[dtype] Verified BF16 execution on CPU -> torch.bfloat16")
return torch.bfloat16
except Exception as e:
logger.warning(f"[dtype] BF16 probe failed: {e}")
logger.info("[dtype] fallback -> torch.float32")
return torch.float32
# Log CPU dtype capability at startup
CPU_DTYPE = _pick_cpu_dtype()
logger.info(f"[init] Default CPU dtype = {CPU_DTYPE}")
_MODEL_CACHE: Dict[tuple[str, torch.dtype], AutoModelForCausalLM] = {}
def _get_model(device: str, dtype: torch.dtype) -> Tuple[AutoModelForCausalLM, torch.dtype]:
key = (device, dtype)
if key in _MODEL_CACHE:
logger.info(f"[cache] hit model for device={device} dtype={dtype}")
return _MODEL_CACHE[key], dtype
logger.info(f"[load] begin from_pretrained device={device} dtype={dtype}")
cfg = AutoConfig.from_pretrained(MODEL_ID, trust_remote_code=True)
if hasattr(cfg, "quantization_config"):
logger.warning("[load] removing quantization_config from config to avoid FP8 path")
delattr(cfg, "quantization_config")
eff_dtype = dtype
try:
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
config=cfg,
torch_dtype=dtype,
trust_remote_code=True,
device_map="auto" if device != "cpu" else {"": "cpu"},
low_cpu_mem_usage=False,
)
except Exception as e:
if device == "cpu" and dtype == torch.bfloat16:
logger.warning(f"[load] BF16 load failed on CPU ({e}). retry FP32.")
eff_dtype = torch.float32
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
config=cfg,
torch_dtype=eff_dtype,
trust_remote_code=True,
device_map={"": "cpu"},
low_cpu_mem_usage=False,
)
else:
logger.exception("[load] from_pretrained failed")
raise
if device == "cpu":
logger.info(f"[load] casting all weights to CPU dtype={eff_dtype}")
model = model.to(device=device, dtype=eff_dtype)
else:
logger.info(f"[load] moving model to device={device} (no recast)")
model = model.to(device=device)
model.eval()
try:
first_dtype = next(model.parameters()).dtype
logger.info(f"[load] ready. effective_dtype={eff_dtype} first_param_dtype={first_dtype}")
except Exception:
logger.info(f"[load] ready. effective_dtype={eff_dtype} (param dtype probe failed)")
_MODEL_CACHE[(device, eff_dtype)] = model
return model, eff_dtype
def _max_context(model, tokenizer) -> int:
mc = getattr(getattr(model, "config", None), "max_position_embeddings", None)
if isinstance(mc, int) and mc > 0:
return mc
tk = getattr(tokenizer, "model_max_length", None)
if isinstance(tk, int) and tk > 0 and tk < 10**12:
return tk
return 32768 # safe default for Qwen3
def _build_inputs_with_truncation(prompt: str, device: str, max_new_tokens: int, model, tokenizer):
toks = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
input_ids = toks["input_ids"]
attn = toks.get("attention_mask", None)
ctx = _max_context(model, tokenizer)
limit = max(8, ctx - max_new_tokens)
in_len = input_ids.shape[-1]
if in_len > limit:
cut = in_len - limit
input_ids = input_ids[:, -limit:]
if attn is not None:
attn = attn[:, -limit:]
logger.warning(f"[truncate] prompt_tokens={in_len} > limit={limit}. truncated_left_by={cut} to fit ctx={ctx}, new_input={input_ids.shape[-1]}, max_new={max_new_tokens}")
inputs = {"input_ids": input_ids}
if attn is not None:
inputs["attention_mask"] = attn
inputs = {k: v.to(device) if hasattr(v, "to") else v for k, v in inputs.items()}
return inputs, in_len, ctx, limit
class HFChatBackend(ChatBackend):
async def stream(self, request: Dict[str, Any]) -> AsyncIterable[Dict[str, Any]]:
if load_error:
raise RuntimeError(load_error)
messages = request.get("messages", [])
tools = request.get("tools")
temperature = float(request.get("temperature", settings.LlmTemp or 0.3))
req_max_tokens = int(request.get("max_tokens", settings.LlmOpenAICtxSize or 32000))
rid = f"chatcmpl-hf-{int(time.time())}"
now = int(time.time())
logger.info(f"[req] rid={rid} temp={temperature} req_max_tokens={req_max_tokens} "
f"msgs={len(messages)} tools={'yes' if tools else 'no'} "
f"spaces={'yes' if spaces else 'no'} cuda={'yes' if torch.cuda.is_available() else 'no'}")
x_ip_token = request.get("x_ip_token")
if x_ip_token and zero_client:
zero_client.HEADERS["X-IP-Token"] = x_ip_token
logger.info("[req] injected X-IP-Token into ZeroGPU headers")
if hasattr(tokenizer, "apply_chat_template") and getattr(tokenizer, "chat_template", None):
try:
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
logger.info(f"[prompt] built via chat_template. len={len(prompt)}\n{_snippet(prompt, 800)}")
except Exception as e:
logger.warning(f"[prompt] chat_template failed -> fallback. err={e}")
prompt = messages[-1]["content"] if messages else "(empty)"
logger.info(f"[prompt] fallback content len={len(prompt)}\n{_snippet(prompt, 800)}")
else:
prompt = messages[-1]["content"] if messages else "(empty)"
logger.info(f"[prompt] no template. using last user text len={len(prompt)}\n{_snippet(prompt, 800)}")
def _run_once(prompt: str, device: str, req_dtype: torch.dtype) -> str:
model, eff_dtype = _get_model(device, req_dtype)
max_new_tokens = req_max_tokens
inputs, orig_in_len, ctx, limit = _build_inputs_with_truncation(prompt, device, max_new_tokens, model, tokenizer)
logger.info(f"[gen] device={device} dtype={eff_dtype} input_tokens={inputs['input_ids'].shape[-1]} "
f"(orig={orig_in_len}) max_ctx={ctx} limit_for_input={limit} max_new_tokens={max_new_tokens}")
do_sample = temperature > 1e-6
temp = max(1e-5, temperature) if do_sample else 0.0
eos_id = tokenizer.eos_token_id
pad_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else eos_id
with torch.inference_mode():
if device != "cpu":
autocast_ctx = torch.autocast(device_type=device, dtype=eff_dtype)
else:
autocast_ctx = torch.cpu.amp.autocast(dtype=torch.bfloat16) if eff_dtype == torch.bfloat16 else nullcontext()
gen_kwargs = dict(
max_new_tokens=max_new_tokens,
temperature=temp,
do_sample=do_sample,
use_cache=True,
eos_token_id=eos_id,
pad_token_id=pad_id,
)
logger.info(f"[gen] kwargs={gen_kwargs}")
with autocast_ctx:
outputs = model.generate(**inputs, **gen_kwargs)
input_len = inputs["input_ids"].shape[-1]
generated_ids = outputs[0][input_len:]
logger.info(f"[gen] new_tokens={generated_ids.shape[-1]}")
text = tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
logger.info(f"[gen] text len={len(text)}\n{_snippet(text, 1200)}")
return text
if spaces:
@spaces.GPU(duration=120)
def run_once_sync(prompt: str) -> str:
if torch.cuda.is_available():
logger.info("[path] ZeroGPU + CUDA")
return _run_once(prompt, device="cuda", req_dtype=torch.float16)
logger.info("[path] ZeroGPU but no CUDA -> CPU fallback")
return _run_once(prompt, device="cpu", req_dtype=_pick_cpu_dtype())
text = await asyncio.to_thread(run_once_sync, prompt)
else:
logger.info("[path] CPU-only runtime")
text = await asyncio.to_thread(_run_once, prompt, "cpu", _pick_cpu_dtype())
chunk = {
"id": rid,
"object": "chat.completion.chunk",
"created": now,
"model": MODEL_ID,
"choices": [
{"index": 0, "delta": {"role": "assistant", "content": text}, "finish_reason": "stop"}
],
}
logger.info(f"[out] chunk summary -> id={rid} content_len={len(text)}")
yield chunk
class StubImagesBackend(ImagesBackend):
async def generate_b64(self, request: Dict[str, Any]) -> str:
logger.warning("Image generation not supported in HF backend.")
return "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR4nGP4BwQACfsD/etCJH0AAAAASUVORK5CYII="