Alikestocode's picture
Add GPU estimator, DDG search, and cancel support
4ce42e8
from __future__ import annotations
import json
import math
import os
import re
import threading
from itertools import islice
from typing import Any, Dict, List, Tuple, Optional
import gradio as gr
import spaces
import torch
from transformers import (
AutoTokenizer,
TextIteratorStreamer,
pipeline,
StoppingCriteria,
StoppingCriteriaList,
)
from threading import Thread
from concurrent.futures import ThreadPoolExecutor
try:
from huggingface_hub import snapshot_download
HF_HUB_AVAILABLE = True
except ImportError: # pragma: no cover
HF_HUB_AVAILABLE = False
try:
from ddgs import DDGS
DDGS_AVAILABLE = True
except ImportError:
DDGS_AVAILABLE = False
# Enable optimizations
torch.backends.cuda.matmul.allow_tf32 = True
# ZeroGPU often exposes MIG UUIDs; keep them unless the variable is empty
MIG_VISIBLE = False
if torch.cuda.is_available():
cuda_visible = os.environ.get("CUDA_VISIBLE_DEVICES", "")
if not cuda_visible:
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
cuda_visible = "0"
print("CUDA_VISIBLE_DEVICES was empty -> set to 0")
elif cuda_visible.startswith("MIG"):
MIG_VISIBLE = True
print(f"CUDA detected: {torch.cuda.get_device_name(0)}")
print(f"CUDA_VISIBLE_DEVICES: {cuda_visible or os.environ.get('CUDA_VISIBLE_DEVICES', 'not set')}")
else:
print("WARNING: CUDA not available - vLLM will not work")
# Try to import vLLM (primary inference engine)
try:
from vllm import LLM, SamplingParams
from vllm.engine.arg_utils import AsyncEngineArgs
VLLM_AVAILABLE = True
except ImportError:
VLLM_AVAILABLE = False
LLM = None
SamplingParams = None
print("Warning: vLLM not available, falling back to Transformers")
cancel_event = threading.Event()
# Optional flag to disable vLLM (defaults to true on MIG due to device detection instability)
DISABLE_VLLM = os.environ.get("DISABLE_VLLM", "1" if MIG_VISIBLE else "0") == "1"
# ---------------------------------------------------------------------------
# Parallel prefetch of model weights/tokenizers to reduce first-load latency
# ---------------------------------------------------------------------------
PREFETCH_DISABLED = os.environ.get("DISABLE_PREFETCH", "0") == "1"
PREFETCH_THREADS = int(os.environ.get("PREFETCH_THREADS", "4"))
PREFETCH_EXECUTOR = None
LOCAL_REPO_CACHE: Dict[str, str] = {}
def _prefetch_repo(repo_id: str) -> None:
if not HF_HUB_AVAILABLE:
return
try:
snapshot_download(
repo_id=repo_id,
etag_timeout=10,
resume_download=True,
local_files_only=False,
)
print(f"Prefetched repo: {repo_id}")
except Exception as exc: # pragma: no cover
print(f"Prefetch skipped for {repo_id}: {exc}")
def _ensure_local_repo(repo_id: str) -> Optional[str]:
if not HF_HUB_AVAILABLE:
return None
cached = LOCAL_REPO_CACHE.get(repo_id)
if cached and os.path.isdir(cached):
return cached
try:
local_path = snapshot_download(
repo_id=repo_id,
etag_timeout=10,
resume_download=True,
local_files_only=False,
)
LOCAL_REPO_CACHE[repo_id] = local_path
return local_path
except Exception as exc: # pragma: no cover
print(f"Local snapshot failed for {repo_id}: {exc}")
return None
def _retrieve_search_results(query: str, max_results: int, max_chars: int) -> List[str]:
if not DDGS_AVAILABLE:
return []
results: List[str] = []
try:
with DDGS() as ddgs:
for idx, item in enumerate(
islice(
ddgs.text(
query,
region="wt-wt",
safesearch="moderate",
timelimit="y",
),
max_results,
)
):
title = (item.get("title") or "Untitled").strip()
body = (item.get("body") or "").strip()
url = (item.get("href") or "").strip()
snippet = body[: max_chars].replace("\n", " ")
formatted = f"[{idx+1}] {title}{snippet}"
if url:
formatted += f" ({url})"
results.append(formatted)
except Exception as exc: # pragma: no cover
print(f"[DEBUG] DDG search failed: {exc}")
return results
class CancelStoppingCriteria(StoppingCriteria):
def __call__(self, input_ids, scores, **kwargs) -> bool:
return cancel_event.is_set()
def estimate_gpu_seconds(
model_name: str,
max_new_tokens: int,
enable_search: bool,
) -> float:
params_b = MODELS.get(model_name, {}).get("params_b", 4.0)
base = 12.0 + params_b * 3.0
tokens_per_sec = max(40.0, 320.0 / (1.0 + params_b / 6.0))
generation_time = max_new_tokens / tokens_per_sec
search_time = 8.0 if enable_search else 0.0
return base + generation_time + search_time
def format_gpu_estimate_message(
model_name: str,
max_new_tokens: int,
enable_search: bool,
) -> Tuple[str, int]:
est_seconds = estimate_gpu_seconds(model_name, max_new_tokens, enable_search)
rounded = int(math.ceil(est_seconds))
recommended = int(math.ceil(max(60, rounded) / 60.0) * 60)
recommended = max(60, min(1800, recommended))
model_size = MODELS.get(model_name, {}).get("params_b", 4.0)
message = (
f"⏱️ **Estimated GPU Time:** ~{rounded} seconds\n\n"
f"📊 **Model Size:** {model_size:.1f}B parameters\n"
f"🔍 **Web Search:** {'Enabled' if enable_search else 'Disabled'}\n"
f"✅ **Suggested GPU Duration slider:** {recommended} seconds"
)
return message, recommended
def update_gpu_controls(
model_name: str,
max_new_tokens: int,
enable_search: bool,
current_duration: int,
):
message, recommended = format_gpu_estimate_message(
model_name,
max_new_tokens,
enable_search,
)
updated_value = current_duration if current_duration >= recommended else recommended
return message, gr.update(value=updated_value)
def _start_prefetch_workers(model_names: list[str]):
global PREFETCH_EXECUTOR
if PREFETCH_DISABLED or not HF_HUB_AVAILABLE:
return
if PREFETCH_EXECUTOR is not None:
return
if not model_names:
return
worker_count = max(1, min(PREFETCH_THREADS, len(model_names) * 2))
PREFETCH_EXECUTOR = ThreadPoolExecutor(max_workers=worker_count, thread_name_prefix="prefetch")
submitted = set()
for model_name in model_names:
repos = {MODELS[model_name]["repo_id"]}
tokenizer_repo = MODELS[model_name].get("tokenizer_repo")
if tokenizer_repo:
repos.add(tokenizer_repo)
for repo in repos:
if repo in submitted:
continue
submitted.add(repo)
PREFETCH_EXECUTOR.submit(_prefetch_repo, repo)
MODELS = {
"Router-Gemma3-27B-AWQ": {
"repo_id": "Alovestocode/router-gemma3-merged-awq", # AWQ quantized model
"tokenizer_repo": "Alovestocode/router-gemma3-merged", # Tokenizer from original repo
"description": "Router checkpoint on Gemma3 27B merged, optimized with AWQ quantization via vLLM.",
"params_b": 27.0,
"quantization": "awq", # vLLM will auto-detect AWQ
},
"Router-Qwen3-32B-AWQ": {
"repo_id": "Alovestocode/router-qwen3-32b-merged-awq", # AWQ quantized model
"tokenizer_repo": "Alovestocode/router-qwen3-32b-merged", # Tokenizer from original repo
"description": "Router checkpoint on Qwen3 32B merged, optimized with AWQ quantization via vLLM.",
"params_b": 32.0,
"quantization": "awq", # vLLM will auto-detect AWQ
},
}
DEFAULT_MODEL = os.environ.get("DEFAULT_ROUTER_MODEL", "Router-Gemma3-27B-AWQ")
if DEFAULT_MODEL not in MODELS:
DEFAULT_MODEL = next(iter(MODELS)) if MODELS else None
def _resolve_prefetch_model_names(include_default: bool) -> list[str]:
entries = os.environ.get("ROUTER_PREFETCH_MODELS")
if entries:
names = [item.strip() for item in entries.split(",") if item.strip()]
else:
single = os.environ.get("ROUTER_PREFETCH_MODEL")
names = [single] if single else []
if names == ["ALL"] or names == ["all"]:
names = list(MODELS.keys())
valid = [name for name in names if name in MODELS]
if not valid and include_default and DEFAULT_MODEL:
valid = [DEFAULT_MODEL]
return valid
_start_prefetch_workers(_resolve_prefetch_model_names(include_default=True))
# Try to import LLM Compressor (for quantization - optional, vLLM has native AWQ support)
# Note: llm-compressor is only needed for quantizing models, not for loading pre-quantized AWQ models
# vLLM can load AWQ models natively without llm-compressor
try:
# Try both package names (llm-compressor and llmcompressor)
try:
from llmcompressor import oneshot
# Correct import path: AWQModifier is in modifiers.awq, not modifiers.quantization
from llmcompressor.modifiers.awq import AWQModifier
except ImportError:
# Try alternative package name
import sys
import subprocess
# Package might be named llm-compressor (with hyphen)
try:
import importlib.util
spec = importlib.util.find_spec("llm_compressor")
if spec is None:
raise ImportError("llm-compressor not found")
from llm_compressor import oneshot
from llm_compressor.modifiers.awq import AWQModifier
except ImportError:
raise ImportError("Neither llmcompressor nor llm-compressor found")
LLM_COMPRESSOR_AVAILABLE = True
print("Info: LLM Compressor available (for quantizing models)")
except ImportError:
LLM_COMPRESSOR_AVAILABLE = False
# This is fine - vLLM has native AWQ support, so we don't need llm-compressor for loading
print("Info: LLM Compressor not available (not needed - vLLM has native AWQ support for pre-quantized models)")
# Try to import AWQ (deprecated, but kept for fallback compatibility)
# Note: AutoAWQ is deprecated; vLLM handles AWQ natively via llm-compressor
try:
from awq import AutoAWQForCausalLM
AWQ_AVAILABLE = True
import warnings
warnings.filterwarnings("ignore", category=DeprecationWarning, module="awq")
except ImportError:
AWQ_AVAILABLE = False
print("Info: AutoAWQ not available (using vLLM native AWQ support instead)")
# Always import BitsAndBytesConfig for fallback
try:
from transformers import BitsAndBytesConfig
BITSANDBYTES_AVAILABLE = True
except ImportError:
BITSANDBYTES_AVAILABLE = False
BitsAndBytesConfig = None
print("Warning: BitsAndBytes not available")
# Try to import FlashAttention-2
try:
import flash_attn
FLASH_ATTN_AVAILABLE = True
except ImportError:
FLASH_ATTN_AVAILABLE = False
print("Warning: FlashAttention-2 not available")
HF_TOKEN = os.environ.get("HF_TOKEN")
if not HF_TOKEN:
raise RuntimeError("HF_TOKEN environment variable must be set for private router checkpoints.")
PLAN_END_TOKEN = "<|end_of_plan|>"
STOP_SEQUENCES = [PLAN_END_TOKEN, "</json>", "</JSON>"]
ROUTER_SYSTEM_PROMPT = """You are the Router Agent coordinating Math, Code, and General-Search specialists.\nEmit EXACTLY ONE strict JSON object with keys route_plan, route_rationale, expected_artifacts,\nthinking_outline, handoff_plan, todo_list, difficulty, tags, acceptance_criteria, metrics.\nRules:\n- No markdown/code fences, no natural-language prologues or epilogues.\n- route_plan must be an ordered list of tool invocations such as /math(...), /code(...), /general-search(...).\n- todo_list must map each checklist item to the responsible tool.\n- metrics must include primary and secondary arrays (add optional *_guidance fields when they exist).\n- After the closing brace of the JSON object, immediately append the sentinel <|end_of_plan|>.\nExample output:\n{\n "route_plan": ["/general-search(...)"],\n "route_rationale": "...",\n ...\n}<|end_of_plan|>\nReturn nothing else."""
REQUIRED_KEYS = [
"route_plan",
"route_rationale",
"expected_artifacts",
"thinking_outline",
"handoff_plan",
"todo_list",
"difficulty",
"tags",
"acceptance_criteria",
"metrics",
]
PIPELINES: Dict[str, Any] = {} # For Transformers fallback
VLLM_MODELS: Dict[str, Any] = {} # For vLLM models
TOKENIZER_CACHE: Dict[str, Any] = {}
WARMED_REMAINING = False
TOOL_PATTERN = re.compile(r"^/[a-z0-9_-]+\(.*\)$", re.IGNORECASE)
def get_tokenizer(repo: str, tokenizer_repo: str = None):
"""Get tokenizer, preferring tokenizer_repo if provided (for AWQ models)."""
# Use tokenizer_repo if provided (for AWQ models where tokenizer is in original repo)
actual_repo = tokenizer_repo if tokenizer_repo else repo
tok = TOKENIZER_CACHE.get(actual_repo)
if tok is not None:
return tok
tok = AutoTokenizer.from_pretrained(
actual_repo,
token=HF_TOKEN,
use_fast=True,
trust_remote_code=True
)
tok.padding_side = "left"
tok.truncation_side = "left"
if tok.pad_token_id is None and tok.eos_token_id is not None:
tok.pad_token_id = tok.eos_token_id
TOKENIZER_CACHE[actual_repo] = tok
return tok
def load_vllm_model(model_name: str):
"""Load model with vLLM (supports AWQ natively, continuous batching, PagedAttention)."""
if model_name in VLLM_MODELS:
return VLLM_MODELS[model_name]
model_config = MODELS[model_name]
repo = model_config["repo_id"]
quantization = model_config.get("quantization", None)
# For AWQ models, vLLM should point to repo root (not default/ subfolder)
# If repo is stored with AWQ artifacts inside a default/ directory, fall back to local snapshot
if quantization == "awq":
model_path = repo
local_repo = _ensure_local_repo(repo)
if local_repo:
default_dir = os.path.join(local_repo, "default")
model_path = default_dir if os.path.isdir(default_dir) else local_repo
print(f"Loading {model_path} (local snapshot) with vLLM (AWQ quantization)...")
else:
print(f"Loading {model_path} with vLLM (AWQ quantization, vLLM will find files in default/ via quantization_config.json)...")
else:
model_path = repo
print(f"Loading {model_path} with vLLM (quantization: {quantization})...")
try:
# Detect device explicitly for vLLM
# vLLM needs explicit device configuration on ZeroGPU
if not torch.cuda.is_available():
raise RuntimeError("CUDA not available - vLLM requires GPU. Falling back to Transformers pipeline.")
print(f" → CUDA available: {torch.cuda.get_device_name(0)}")
print(f" → CUDA device count: {torch.cuda.device_count()}")
print(f" → CUDA_VISIBLE_DEVICES: {os.environ.get('CUDA_VISIBLE_DEVICES', 'not set')}")
# vLLM configuration optimized for ZeroGPU H200 slice
# vLLM natively supports AWQ via llm-compressor (replaces deprecated AutoAWQ)
# Note: HF_TOKEN is passed via environment variable, not as a parameter
# vLLM auto-detects CUDA from torch.cuda.is_available() and CUDA_VISIBLE_DEVICES
# For AWQ models with files in default/ subfolder, vLLM should auto-detect via quantization_config.json
llm_kwargs = {
"model": model_path, # Use model_path which may point to default/ subfolder
"trust_remote_code": True,
"dtype": "bfloat16", # Prefer bf16 over int8 for speed
"gpu_memory_utilization": 0.90, # Leave headroom for KV cache
"max_model_len": 16384, # Adjust based on GPU memory
"enable_chunked_prefill": True, # Better for long prompts
"tensor_parallel_size": 1, # Single GPU for ZeroGPU
"max_num_seqs": 256, # Continuous batching capacity
"enable_prefix_caching": True, # Cache prompts for faster TTFT
}
# Ensure CUDA_VISIBLE_DEVICES is set correctly for vLLM device detection
# ZeroGPU exposes MIG UUIDs; keep them unless the variable is empty
cuda_visible = os.environ.get("CUDA_VISIBLE_DEVICES", "")
if not cuda_visible:
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
cuda_visible = "0"
print(" → CUDA_VISIBLE_DEVICES was empty, set to 0")
try:
if hasattr(torch.cuda, '_lazy_init'):
torch.cuda._lazy_init()
except Exception:
pass
else:
print(f" → CUDA_VISIBLE_DEVICES retained: {cuda_visible}")
# Force torch to see the correct device after ensuring CUDA_VISIBLE_DEVICES
if torch.cuda.is_available():
device_name = torch.cuda.get_device_name(0)
print(f" → Verified CUDA device accessible: {device_name}")
torch.cuda.set_device(0)
print(" → Set torch.cuda default device to 0")
# Disable Ray executor on ZeroGPU to simplify device handling
os.environ.setdefault("VLLM_USE_RAY", "0")
os.environ.setdefault("VLLM_WORKER_USE_RAY", "0")
# Add quantization if specified (vLLM auto-detects AWQ via llm-compressor)
if quantization == "awq":
llm_kwargs["quantization"] = "awq"
# AWQ model files are in the 'default' subfolder
# vLLM should auto-detect this via quantization_config.json at repo root
# If auto-detection fails, we can explicitly point to default/ subfolder
# Enable FP8 KV cache for 50% memory reduction (allows longer contexts)
# FP8 KV cache is compatible with AWQ quantization
try:
llm_kwargs["kv_cache_dtype"] = "fp8"
print(f" → AWQ quantization + FP8 KV cache enabled (vLLM native support)")
print(f" → FP8 KV cache reduces memory by ~50%, enabling longer contexts")
print(f" → Loading AWQ model from: {model_path} (files in default/ subfolder)")
except Exception:
# Fallback if FP8 KV cache not supported
print(f" → AWQ quantization enabled (FP8 KV cache not available)")
print(f" → Loading AWQ model from: {model_path} (files in default/ subfolder)")
elif quantization == "fp8":
# Try FP8 quantization if available (faster than AWQ)
try:
llm_kwargs["quantization"] = "fp8"
llm_kwargs["dtype"] = "float8_e5m2"
print(f" → FP8 quantization enabled (~2x faster than AWQ)")
except Exception:
print(f" → FP8 quantization not available, falling back to bf16")
# vLLM will now detect the CUDA device via torch / environment settings above
print(f" → Loading with vLLM (continuous batching, PagedAttention)...")
llm = LLM(**llm_kwargs)
VLLM_MODELS[model_name] = llm
print(f"✅ vLLM model loaded: {model_name}")
print(f" - Continuous batching: enabled (max {llm_kwargs['max_num_seqs']} concurrent)")
print(f" - Prefix caching: enabled")
print(f" - Quantization: {quantization or 'none (bf16)'}")
return llm
except Exception as exc:
print(f"❌ vLLM load failed for {repo}: {exc}")
import traceback
traceback.print_exc()
raise
def load_awq_pipeline(repo: str, tokenizer):
"""Load AWQ-quantized model with FlashAttention-2 and torch.compile (Transformers fallback)."""
model = AutoAWQForCausalLM.from_quantized(
repo,
fuse_layers=True,
trust_remote_code=True,
device_map="auto",
token=HF_TOKEN,
)
# Prepare model kwargs with FlashAttention-2 if available
model_kwargs = {}
if FLASH_ATTN_AVAILABLE:
model_kwargs["attn_implementation"] = "flash_attention_2"
pipe = pipeline(
task="text-generation",
model=model,
tokenizer=tokenizer,
trust_remote_code=True,
device_map="auto",
model_kwargs=model_kwargs,
use_cache=True,
torch_dtype=torch.bfloat16, # Prefer bf16 over int8 for speed
)
pipe.model.eval()
# Apply torch.compile for kernel fusion (~10-20% speedup after first call)
try:
if hasattr(torch, 'compile'):
print("Applying torch.compile for kernel fusion...")
pipe.model = torch.compile(pipe.model, mode="reduce-overhead")
print("✅ torch.compile applied (first call will be slower, subsequent calls faster)")
except Exception as exc:
print(f"⚠️ torch.compile failed: {exc} (continuing without compilation)")
return pipe
def load_pipeline(model_name: str):
"""Load model with vLLM (preferred) or Transformers (fallback).
Fallback chain:
1. vLLM with AWQ (best performance, continuous batching)
2. vLLM with FP16 (if AWQ not available)
3. Transformers with AWQ (via AutoAWQ - deprecated but functional)
4. Transformers with BitsAndBytes 8-bit
5. Transformers with FP16/FP32
"""
# Try vLLM first (best performance with native AWQ support via llm-compressor)
# vLLM handles AWQ natively, so AutoAWQ deprecation doesn't affect us
if VLLM_AVAILABLE and not DISABLE_VLLM:
try:
print(f"🔄 Attempting to load {model_name} with vLLM (native AWQ support)...")
return load_vllm_model(model_name)
except Exception as exc:
print(f"⚠️ vLLM load failed: {exc}")
print(f" → Falling back to Transformers pipeline...")
import traceback
traceback.print_exc()
# Fallback to Transformers pipeline
if model_name in PIPELINES:
print(f"✅ Using cached Transformers pipeline for {model_name}")
return PIPELINES[model_name]
if DISABLE_VLLM and VLLM_AVAILABLE:
print("⚠️ vLLM disabled for this deployment (DISABLE_VLLM=1 or MIG device detected)")
model_config = MODELS[model_name]
repo = model_config["repo_id"]
tokenizer_repo = model_config.get("tokenizer_repo", None)
quantization = model_config.get("quantization", None)
# For AWQ models, the AWQ repo doesn't have standard model files (they're in default/)
# Use the original repo for Transformers fallback, not the AWQ repo
if quantization == "awq" and tokenizer_repo:
# AWQ repos have files in default/ subfolder which Transformers can't load directly
# Use the original repo for Transformers fallback
transformers_repo = tokenizer_repo # Use original repo for Transformers
print(f"⚠️ AWQ model detected - Transformers fallback will use original repo: {transformers_repo}")
else:
transformers_repo = repo
tokenizer = get_tokenizer(repo, tokenizer_repo=tokenizer_repo)
# Try AWQ first if available (Transformers fallback path)
if AWQ_AVAILABLE:
try:
print(f"🔄 Loading {transformers_repo} with Transformers + AutoAWQ (fallback path)...")
pipe = load_awq_pipeline(transformers_repo, tokenizer)
PIPELINES[model_name] = pipe
_schedule_background_warm(model_name)
# Warm kernels immediately after loading
Thread(target=lambda: _warm_kernels(model_name), daemon=True).start()
print(f"✅ Transformers + AutoAWQ pipeline loaded: {model_name}")
return pipe
except Exception as exc:
print(f"⚠️ AutoAWQ load failed for {transformers_repo}: {exc}")
print(f" → Falling back to BitsAndBytes 8-bit...")
# Fallback to BitsAndBytes 8-bit
if BITSANDBYTES_AVAILABLE:
try:
print(f"🔄 Loading {transformers_repo} with BitsAndBytes 8-bit quantization...")
quant_config = BitsAndBytesConfig(load_in_8bit=True)
model_kwargs = {"quantization_config": quant_config}
if FLASH_ATTN_AVAILABLE:
model_kwargs["attn_implementation"] = "flash_attention_2"
pipe = pipeline(
task="text-generation",
model=transformers_repo,
tokenizer=tokenizer,
trust_remote_code=True,
device_map="auto",
model_kwargs=model_kwargs,
use_cache=True,
token=HF_TOKEN,
torch_dtype=torch.bfloat16,
)
pipe.model.eval()
# Apply torch.compile for kernel fusion (~10-20% speedup after first call)
try:
if hasattr(torch, 'compile'):
pipe.model = torch.compile(pipe.model, mode="reduce-overhead")
except Exception:
pass
PIPELINES[model_name] = pipe
_schedule_background_warm(model_name)
print(f"✅ BitsAndBytes 8-bit pipeline loaded: {model_name}")
return pipe
except Exception as exc:
print(f"⚠️ BitsAndBytes 8-bit load failed for {transformers_repo}: {exc}")
print(f" → Falling back to FP16/FP32...")
# Fallback to bfloat16/fp16/fp32 (unquantized)
for dtype in (torch.bfloat16, torch.float16, torch.float32):
dtype_name = {torch.bfloat16: "bfloat16", torch.float16: "float16", torch.float32: "float32"}[dtype]
try:
print(f"🔄 Loading {transformers_repo} with {dtype_name} precision...")
model_kwargs = {}
if FLASH_ATTN_AVAILABLE:
model_kwargs["attn_implementation"] = "flash_attention_2"
pipe = pipeline(
task="text-generation",
model=transformers_repo,
tokenizer=tokenizer,
trust_remote_code=True,
device_map="auto",
dtype=dtype,
model_kwargs=model_kwargs,
use_cache=True,
token=HF_TOKEN,
)
pipe.model.eval()
# Apply torch.compile for kernel fusion
try:
if hasattr(torch, 'compile'):
pipe.model = torch.compile(pipe.model, mode="reduce-overhead")
except Exception:
pass
PIPELINES[model_name] = pipe
_schedule_background_warm(model_name)
print(f"✅ {dtype_name} pipeline loaded: {model_name}")
return pipe
except Exception as exc:
print(f"⚠️ {dtype_name} load failed: {exc}")
continue
# Final fallback (no quantization, no FlashAttention)
print(f"⚠️ All quantization methods failed, using basic pipeline...")
model_kwargs = {}
if FLASH_ATTN_AVAILABLE:
model_kwargs["attn_implementation"] = "flash_attention_2"
pipe = pipeline(
task="text-generation",
model=transformers_repo,
tokenizer=tokenizer,
trust_remote_code=True,
device_map="auto",
model_kwargs=model_kwargs,
use_cache=True,
token=HF_TOKEN,
)
pipe.model.eval()
# Apply torch.compile for kernel fusion
try:
if hasattr(torch, 'compile'):
pipe.model = torch.compile(pipe.model, mode="reduce-overhead")
except Exception:
pass
PIPELINES[model_name] = pipe
_schedule_background_warm(model_name)
print(f"✅ Basic pipeline loaded: {model_name}")
return pipe
def _warm_kernels(model_name: str) -> None:
"""Warm up CUDA kernels with a small dummy generation."""
try:
# Check if using vLLM
if VLLM_AVAILABLE and model_name in VLLM_MODELS:
llm = VLLM_MODELS[model_name]
# vLLM handles warmup internally, but we can trigger a small generation
sampling_params = SamplingParams(temperature=0.0, max_tokens=2)
_ = llm.generate("test", sampling_params)
print(f"vLLM kernels warmed for {model_name}")
return
# Transformers pipeline warmup
pipe = PIPELINES.get(model_name)
if pipe is None:
return
tokenizer = pipe.tokenizer
# Create a minimal prompt for warmup
warmup_text = "test"
inputs = tokenizer(warmup_text, return_tensors="pt")
if hasattr(pipe.model, 'device'):
inputs = {k: v.to(pipe.model.device) for k, v in inputs.items()}
elif torch.cuda.is_available():
inputs = {k: v.cuda() for k, v in inputs.items()}
# Run a tiny generation to JIT-fuse kernels
with torch.inference_mode():
_ = pipe.model.generate(
**inputs,
max_new_tokens=2,
do_sample=False,
use_cache=True,
)
print(f"Transformers kernels warmed for {model_name}")
except Exception as exc:
print(f"Kernel warmup failed for {model_name}: {exc}")
def _schedule_background_warm(loaded_model: str) -> None:
global WARMED_REMAINING
if WARMED_REMAINING:
return
warm_remaining = os.environ.get("ROUTER_WARM_REMAINING", "1")
if warm_remaining not in {"1", "true", "True"}:
return
# Check both PIPELINES and VLLM_MODELS for remaining models
loaded_models = set(PIPELINES.keys()) | set(VLLM_MODELS.keys())
remaining = [name for name in MODELS if name not in loaded_models]
if not remaining:
WARMED_REMAINING = True
return
def _warm_all():
for name in remaining:
try:
print(f"Background warm start for {name}")
load_pipeline(name)
# Warm kernels after loading
_warm_kernels(name)
except Exception as exc: # pragma: no cover
print(f"Warm start failed for {name}: {exc}")
WARMED_REMAINING = True
Thread(target=_warm_all, daemon=True).start()
def build_router_prompt(
user_task: str,
context: str,
acceptance: str,
extra_guidance: str,
difficulty: str,
tags: str,
) -> str:
prompt_parts = [ROUTER_SYSTEM_PROMPT.strip(), "\n### Router Inputs\n"]
prompt_parts.append(f"Difficulty: {difficulty or 'intermediate'}")
prompt_parts.append(f"Tags: {tags or 'general'}")
if acceptance.strip():
prompt_parts.append(f"Acceptance criteria: {acceptance.strip()}")
if extra_guidance.strip():
prompt_parts.append(f"Additional guidance: {extra_guidance.strip()}")
if context.strip():
prompt_parts.append("\n### Supporting context\n" + context.strip())
prompt_parts.append("\n### User task\n" + user_task.strip())
prompt_parts.append("\nReturn only JSON.")
return "\n".join(prompt_parts)
def extract_json_from_text(text: str) -> str:
start = text.find("{")
if start == -1:
raise ValueError("Router output did not contain a JSON object.")
depth = 0
in_string = False
escape = False
for idx in range(start, len(text)):
ch = text[idx]
if in_string:
if escape:
escape = False
elif ch == "\\":
escape = True
elif ch == '"':
in_string = False
continue
if ch == '"':
in_string = True
continue
if ch == '{':
depth += 1
elif ch == '}':
depth -= 1
if depth == 0:
return text[start : idx + 1]
raise ValueError("Router output JSON appears truncated.")
def trim_at_stop_sequences(text: str) -> Tuple[str, bool]:
"""Trim text at stop sequences and return trimmed text and whether a stop was found."""
earliest = None
for stop in STOP_SEQUENCES:
idx = text.find(stop)
if idx != -1 and (earliest is None or idx < earliest):
earliest = idx
if earliest is not None:
return text[:earliest], True
return text, False
def is_function_call(text: str) -> bool:
return bool(TOOL_PATTERN.match(text.strip()))
def validate_router_plan(plan: Dict[str, Any]) -> Tuple[bool, List[str]]:
issues: List[str] = []
for key in REQUIRED_KEYS:
if key not in plan:
issues.append(f"Missing key: {key}")
route_plan = plan.get("route_plan")
if isinstance(route_plan, str) and is_function_call(route_plan):
plan["route_plan"] = [route_plan]
route_plan = plan["route_plan"]
if not isinstance(route_plan, list) or not route_plan:
issues.append("route_plan must be a non-empty list of tool calls")
else:
cleaned: List[str] = []
for entry in route_plan:
if isinstance(entry, str) and is_function_call(entry.strip().strip("'\"")):
cleaned.append(entry.strip().strip("'\""))
else:
issues.append(f"route_plan entry is not a tool call: {entry}")
if cleaned:
plan["route_plan"] = cleaned
metrics = plan.get("metrics")
if not isinstance(metrics, dict):
issues.append("metrics must be an object containing primary/secondary entries")
todo = plan.get("todo_list")
if not isinstance(todo, list) or not todo:
issues.append("todo_list must contain at least one checklist item")
else:
cleaned_todo: List[str] = []
for entry in todo:
if isinstance(entry, str):
text = entry.strip()
if not text.startswith("- ["):
text = text.lstrip("- ")
text = f"- [ ] {text}"
cleaned_todo.append(text)
else:
issues.append("todo_list entry must be a string")
if cleaned_todo:
plan["todo_list"] = cleaned_todo
return len(issues) == 0, issues
def format_validation_message(ok: bool, issues: List[str]) -> str:
if ok:
return "✅ Router plan includes all required fields."
bullets = "\n".join(f"- {issue}" for issue in issues)
return f"❌ Issues detected:\n{bullets}"
def _generate_router_plan_streaming_internal(
user_task: str,
context: str,
acceptance: str,
extra_guidance: str,
difficulty: str,
tags: str,
model_choice: str,
max_new_tokens: int,
temperature: float,
top_p: float,
gpu_duration: int,
enable_search: bool,
search_max_results: int,
search_max_chars: int,
search_timeout: float,
):
"""Internal generator function for streaming token output."""
if not user_task.strip():
yield "", {}, "❌ User task is required.", ""
return
if model_choice not in MODELS:
yield "", {}, f"❌ Invalid model choice: {model_choice}. Available: {list(MODELS.keys())}", ""
return
cancel_event.clear()
cancelled = False
try:
search_snippets: List[str] = []
if enable_search and DDGS_AVAILABLE and user_task.strip():
search_snippets_holder: List[str] = []
search_error: Optional[Exception] = None
def _fetch_search():
nonlocal search_error
try:
results = _retrieve_search_results(
user_task,
max(1, int(search_max_results)),
max(30, int(search_max_chars)),
)
search_snippets_holder.extend(results)
except Exception as exc: # pragma: no cover
search_error = exc
search_thread = Thread(target=_fetch_search, daemon=True)
search_thread.start()
search_thread.join(timeout=float(max(0.5, search_timeout)))
if search_thread.is_alive():
print("[DEBUG] Search thread timed out; continuing without results.")
if search_error:
print(f"[DEBUG] Search error: {search_error}")
search_snippets = search_snippets_holder
context_for_prompt = context
if search_snippets:
search_block = "\n".join(f"- {snippet}" for snippet in search_snippets)
addendum = (
"\n\n# Web Search Findings\n"
"Use the following snippets as supplementary evidence. "
"Cite them as needed in the generated plan.\n"
f"{search_block}"
)
context_for_prompt = (context_for_prompt or "").rstrip() + addendum
prompt = build_router_prompt(
user_task=user_task,
context=context_for_prompt,
acceptance=acceptance,
extra_guidance=extra_guidance,
difficulty=difficulty,
tags=tags,
)
print(f"[DEBUG] Loading model: {model_choice}")
generator = load_pipeline(model_choice)
print(f"[DEBUG] Model loaded successfully: {type(generator)}")
# Check if using vLLM or Transformers
is_vllm = VLLM_AVAILABLE and isinstance(generator, LLM)
if is_vllm:
# Use vLLM streaming API with continuous batching
# Optimized sampling parameters for router plan generation
sampling_params = SamplingParams(
temperature=temperature,
top_p=top_p,
max_tokens=max_new_tokens,
stop=STOP_SEQUENCES,
skip_special_tokens=False, # Keep special tokens for parsing
spaces_between_special_tokens=False, # Don't add spaces around special tokens
include_stop_str_in_output=False, # Don't include stop sequences in output
)
# vLLM streaming generation (non-blocking, continuous batching)
completion = ""
parsed_plan: Dict[str, Any] | None = None
validation_msg = "🔄 Generating..."
# vLLM's generate with stream=True returns RequestOutput iterator
# Each RequestOutput contains incremental text updates
stream = generator.generate(prompt, sampling_params, stream=True)
prev_text_len = 0
for request_output in stream:
if cancel_event.is_set():
cancelled = True
try:
if hasattr(generator, "abort_request"):
generator.abort_request(request_output.request_id)
except Exception:
pass
break
if not request_output.outputs:
continue
# Get the latest output (vLLM provides incremental updates)
output = request_output.outputs[0]
current_text = output.text
# Extract only new tokens since last update
if len(current_text) > prev_text_len:
new_text = current_text[prev_text_len:]
completion += new_text
prev_text_len = len(current_text)
chunk = completion
finished = False
display_plan = parsed_plan or {}
chunk, finished = trim_at_stop_sequences(chunk)
try:
json_block = extract_json_from_text(chunk)
candidate_plan = json.loads(json_block)
ok, issues = validate_router_plan(candidate_plan)
validation_msg = format_validation_message(ok, issues)
parsed_plan = candidate_plan if ok else parsed_plan
display_plan = candidate_plan
except Exception:
# Ignore until JSON is complete
pass
yield chunk, display_plan, validation_msg, prompt
if finished:
completion = chunk
break
# Check if generation is finished
if request_output.finished:
break
else:
# Use Transformers pipeline (fallback)
# Get the underlying model and tokenizer
model = generator.model
tokenizer = generator.tokenizer
# Set up streaming
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
# Prepare inputs
inputs = tokenizer(prompt, return_tensors="pt")
if hasattr(model, 'device'):
inputs = {k: v.to(model.device) for k, v in inputs.items()}
elif torch.cuda.is_available():
inputs = {k: v.cuda() for k, v in inputs.items()}
# Start generation in a separate thread
generation_kwargs = {
**inputs,
"max_new_tokens": max_new_tokens,
"temperature": temperature,
"top_p": top_p,
"do_sample": True,
"streamer": streamer,
"eos_token_id": tokenizer.eos_token_id,
"pad_token_id": tokenizer.pad_token_id or tokenizer.eos_token_id,
"stopping_criteria": StoppingCriteriaList([CancelStoppingCriteria()]),
}
generation_error = None
def _generate():
nonlocal generation_error
try:
with torch.inference_mode():
model.generate(**generation_kwargs)
except Exception as e:
generation_error = e
print(f"[DEBUG] Generation thread error: {e}")
import traceback
traceback.print_exc()
thread = Thread(target=_generate)
thread.start()
# Stream tokens
completion = ""
parsed_plan: Dict[str, Any] | None = None
validation_msg = "🔄 Generating..."
print(f"[DEBUG] Starting to consume streamer...")
token_count = 0
try:
for new_text in streamer:
if cancel_event.is_set():
cancelled = True
break
if generation_error:
raise generation_error
if new_text:
token_count += 1
completion += new_text
chunk = completion
finished = False
display_plan = parsed_plan or {}
chunk, finished = trim_at_stop_sequences(chunk)
try:
json_block = extract_json_from_text(chunk)
candidate_plan = json.loads(json_block)
ok, issues = validate_router_plan(candidate_plan)
validation_msg = format_validation_message(ok, issues)
parsed_plan = candidate_plan if ok else parsed_plan
display_plan = candidate_plan
except Exception:
# Ignore until JSON is complete
pass
yield chunk, display_plan, validation_msg, prompt
if finished:
completion = chunk
break
print(f"[DEBUG] Streamer finished. Received {token_count} tokens.")
except Exception as stream_error:
print(f"[DEBUG] Streamer error: {stream_error}")
import traceback
traceback.print_exc()
# Wait for thread to finish
thread.join(timeout=5.0)
if generation_error:
raise generation_error
raise stream_error
# Final processing after streaming completes
thread.join(timeout=30.0)
if thread.is_alive():
print("[DEBUG] WARNING: Generation thread still running after timeout")
if generation_error:
raise generation_error
completion = trim_at_stop_sequences(completion.strip())[0]
print(f"[DEBUG] Final completion length: {len(completion)}")
if cancelled:
validation_msg = "⏹️ Generation cancelled by user."
elif not completion:
print("[DEBUG] WARNING: Completion is empty - model may not have generated output")
validation_msg = "⚠️ Model generated empty output. Check GPU allocation and model loading."
elif parsed_plan is None:
try:
json_block = extract_json_from_text(completion)
parsed_plan = json.loads(json_block)
ok, issues = validate_router_plan(parsed_plan)
validation_msg = format_validation_message(ok, issues)
except Exception as exc:
parsed_plan = {}
validation_msg = f"❌ JSON parsing failed: {exc}"
print(f"[DEBUG] JSON parsing error: {exc}")
yield completion, parsed_plan, validation_msg, prompt
except Exception as exc:
import traceback
print(f"[DEBUG] Exception in generation: {exc}")
print(f"[DEBUG] Traceback: {traceback.format_exc()}")
error_msg = f"❌ Generation failed: {str(exc)}"
yield "", {}, error_msg, ""
# Pre-create GPU wrappers for common durations at module load time
# This ensures spaces.GPU decorators are detected during startup
_GPU_WRAPPERS: Dict[int, Any] = {}
# Create wrappers for durations: 60, 120, 180, 240, 300, 360, 420, 480, 540, 600,
# 720, 840, 960, 1080, 1200, 1320, 1440, 1560, 1680, 1800 (every 60s from 60 to 1800)
def _make_gpu_wrapper(duration: int):
"""Factory function to create GPU-decorated wrapper with closure over duration."""
@spaces.GPU(duration=duration)
def wrapper(
user_task: str,
context: str,
acceptance: str,
extra_guidance: str,
difficulty: str,
tags: str,
model_choice: str,
max_new_tokens: int,
temperature: float,
top_p: float,
gpu_duration: int,
enable_search: bool,
search_max_results: int,
search_max_chars: int,
search_timeout: float,
):
yield from _generate_router_plan_streaming_internal(
user_task, context, acceptance, extra_guidance,
difficulty, tags, model_choice, max_new_tokens,
temperature, top_p, duration,
enable_search,
search_max_results,
search_max_chars,
search_timeout,
)
return wrapper
# Pre-create all wrappers at module load time
for duration in range(60, 1801, 60):
_GPU_WRAPPERS[duration] = _make_gpu_wrapper(duration)
def generate_router_plan_streaming(
user_task: str,
context: str,
acceptance: str,
extra_guidance: str,
difficulty: str,
tags: str,
model_choice: str,
max_new_tokens: int,
temperature: float,
top_p: float,
gpu_duration: int = 600,
enable_search: bool = False,
search_max_results: int = 4,
search_max_chars: int = 120,
search_timeout: float = 5.0,
):
"""
Generate router plan with streaming output.
Uses user-specified gpu_duration to select the appropriate GPU wrapper.
"""
# Round to nearest 60 seconds and clamp between 60 and 1800
rounded_duration = ((gpu_duration + 30) // 60) * 60
rounded_duration = max(60, min(1800, rounded_duration))
# Get the pre-created wrapper with this duration
wrapper = _GPU_WRAPPERS[rounded_duration]
yield from wrapper(
user_task, context, acceptance, extra_guidance,
difficulty, tags, model_choice, max_new_tokens,
temperature, top_p, rounded_duration,
enable_search,
int(search_max_results),
int(search_max_chars),
float(search_timeout),
)
def clear_outputs():
return "", {}, "Awaiting generation.", ""
def cancel_generation():
cancel_event.set()
return "⏹️ Cancel request sent. Finishing current step..."
def build_ui():
description = "Use the CourseGPT-Pro router checkpoints (Gemma3/Qwen3) hosted on ZeroGPU to generate structured routing plans."
initial_estimate_text, initial_recommended_duration = format_gpu_estimate_message(
DEFAULT_MODEL,
16000,
False,
)
with gr.Blocks(theme=gr.themes.Soft(), css="""
textarea { font-family: 'JetBrains Mono', 'Fira Code', monospace; }
.status-ok { color: #0d9488; font-weight: 600; }
.status-bad { color: #dc2626; font-weight: 600; }
""") as demo:
gr.Markdown("# 🛰️ Router Control Room — ZeroGPU" )
gr.Markdown(description)
with gr.Row():
with gr.Column(scale=3):
user_task = gr.Textbox(
label="User Task / Problem Statement",
placeholder="Describe the homework-style query that needs routing...",
lines=8,
value="Explain how to solve a constrained optimization homework problem that mixes calculus and coding steps.",
)
context = gr.Textbox(
label="Supporting Context (optional)",
placeholder="Paste any retrieved evidence, PDFs, or rubric notes.",
lines=4,
)
acceptance = gr.Textbox(
label="Acceptance Criteria",
placeholder="Bullet list of 'definition of done' checks.",
lines=3,
value="- Provide citations for every claim.\n- Ensure /math verifies /code output.",
)
extra_guidance = gr.Textbox(
label="Additional Guidance",
placeholder="Special constraints, tools to avoid, etc.",
lines=3,
)
with gr.Column(scale=2):
model_choice = gr.Dropdown(
label="Router Checkpoint",
choices=list(MODELS.keys()),
value=DEFAULT_MODEL,
allow_custom_value=False,
)
difficulty = gr.Radio(
label="Difficulty Tier",
choices=["introductory", "intermediate", "advanced"],
value="advanced",
interactive=True,
)
tags = gr.Textbox(
label="Tags",
placeholder="Comma-separated e.g. calculus, optimization, python",
value="calculus, optimization, python",
)
max_new_tokens = gr.Slider(256, 20000, value=16000, step=32, label="Max New Tokens")
temperature = gr.Slider(0.0, 1.5, value=0.2, step=0.05, label="Temperature")
top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
enable_search = gr.Checkbox(
label="Enable DuckDuckGo Web Search",
value=False,
interactive=DDGS_AVAILABLE,
info="Augment context with live snippets." if DDGS_AVAILABLE else "Install 'ddgs' package to enable search.",
)
with gr.Accordion("Web Search Settings", open=False, visible=DDGS_AVAILABLE) as search_settings:
search_max_results = gr.Slider(
minimum=1,
maximum=10,
value=4,
step=1,
label="Search Results",
interactive=DDGS_AVAILABLE,
)
search_max_chars = gr.Slider(
minimum=50,
maximum=400,
value=160,
step=10,
label="Max Characters per Result",
interactive=DDGS_AVAILABLE,
)
search_timeout = gr.Slider(
minimum=1.0,
maximum=20.0,
value=5.0,
step=0.5,
label="Search Timeout (seconds)",
interactive=DDGS_AVAILABLE,
)
gpu_estimate_display = gr.Markdown(
value=initial_estimate_text,
elem_classes="status-ok",
)
gpu_duration = gr.Slider(
60,
1800,
value=initial_recommended_duration,
step=60,
label="GPU Duration (seconds)",
info="Maximum GPU time allocation for this request",
)
with gr.Row():
generate_btn = gr.Button("Generate Router Plan", variant="primary", scale=1)
clear_btn = gr.Button("Clear", variant="secondary", scale=1)
cancel_btn = gr.Button("Cancel", variant="stop", scale=1)
with gr.Row():
raw_output = gr.Textbox(label="Raw Model Output", lines=12)
plan_json = gr.JSON(label="Parsed Router Plan")
validation_msg = gr.Markdown("Awaiting generation.")
prompt_view = gr.Textbox(label="Full Prompt", lines=10)
generate_btn.click(
generate_router_plan_streaming,
inputs=[
user_task,
context,
acceptance,
extra_guidance,
difficulty,
tags,
model_choice,
max_new_tokens,
temperature,
top_p,
gpu_duration,
enable_search,
search_max_results,
search_max_chars,
search_timeout,
],
outputs=[raw_output, plan_json, validation_msg, prompt_view],
show_progress="full",
api_name="/generate_router_plan_streaming",
)
clear_btn.click(
fn=clear_outputs,
outputs=[raw_output, plan_json, validation_msg, prompt_view],
api_name="/clear_outputs",
)
cancel_btn.click(
fn=cancel_generation,
outputs=[validation_msg],
)
model_choice.change(
fn=update_gpu_controls,
inputs=[model_choice, max_new_tokens, enable_search, gpu_duration],
outputs=[gpu_estimate_display, gpu_duration],
)
max_new_tokens.change(
fn=update_gpu_controls,
inputs=[model_choice, max_new_tokens, enable_search, gpu_duration],
outputs=[gpu_estimate_display, gpu_duration],
)
enable_search.change(
fn=update_gpu_controls,
inputs=[model_choice, max_new_tokens, enable_search, gpu_duration],
outputs=[gpu_estimate_display, gpu_duration],
)
return demo
def _prefetch_from_env() -> None:
names = _resolve_prefetch_model_names(include_default=False)
for name in names:
if name not in MODELS:
print(f"Prefetch skipped, unknown model: {name}")
continue
try:
load_pipeline(name)
print(f"Prefetched router model: {name}")
except Exception as exc: # pragma: no cover
print(f"Prefetch failed for {name}: {exc}")
_prefetch_from_env()
demo = build_ui()
if __name__ == "__main__": # pragma: no cover
# Support both Hugging Face Spaces and Google Cloud Run
# Cloud Run uses PORT, Hugging Face Spaces uses GRADIO_SERVER_PORT
port = int(os.environ.get("PORT", os.environ.get("GRADIO_SERVER_PORT", 7860)))
server_name = os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0")
demo.launch(
server_name=server_name,
server_port=port,
show_api=True
)