ai-model.py with LLMs and embedding models
This commit is contained in:
llm
601
.local/share/pytorch_pod/python-apps/ai-model.py
Executable file
601
.local/share/pytorch_pod/python-apps/ai-model.py
Executable file
@@ -0,0 +1,601 @@
|
||||
#!/usr/bin/env python
|
||||
import base64
|
||||
import gc
|
||||
import io
|
||||
import os
|
||||
import threading
|
||||
import time
|
||||
import uuid
|
||||
from typing import List, Optional, Union, Dict, Any, Literal
|
||||
|
||||
import torch
|
||||
from colpali_engine.models import ColQwen2_5, ColQwen2_5_Processor
|
||||
from fastapi import FastAPI, HTTPException, Request
|
||||
from PIL import Image, ImageFile
|
||||
from pydantic import BaseModel, Field
|
||||
from transformers import (
|
||||
AutoModelForCausalLM,
|
||||
AutoTokenizer,
|
||||
AutoProcessor,
|
||||
AutoModel,
|
||||
)
|
||||
from transformers.utils.import_utils import is_flash_attn_2_available
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
# Configuration
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
# Embedding Models
|
||||
MODEL_ID_NOMIC = "nomic-ai/colnomic-embed-multimodal-7b"
|
||||
MODEL_ID_EVO_7B = "ApsaraStackMaaS/EvoQwen2.5-VL-Retriever-7B-v1"
|
||||
|
||||
# Generation Models
|
||||
MODEL_ID_QWEN3_VL_8B_INSTRUCT = "Qwen/Qwen3-VL-8B-Instruct"
|
||||
MODEL_ID_QWEN3_VL_8B_INSTRUCT_FP8 = "Qwen/Qwen3-VL-8B-Instruct-FP8"
|
||||
MODEL_ID_QWEN3_VL_8B_THINKING = "Qwen/Qwen3-VL-8B-Thinking"
|
||||
MODEL_ID_QWEN3_VL_8B_THINKING_FP8 = "Qwen/Qwen3-VL-8B-Thinking-FP8"
|
||||
MODEL_ID_GPT_OSS_20B = "openai/gpt-oss-20b"
|
||||
|
||||
ALLOWED_EMBEDDING_MODELS = {MODEL_ID_NOMIC, MODEL_ID_EVO_7B}
|
||||
ALLOWED_GENERATION_MODELS = {
|
||||
MODEL_ID_QWEN3_VL_8B_INSTRUCT,
|
||||
MODEL_ID_QWEN3_VL_8B_INSTRUCT_FP8,
|
||||
MODEL_ID_QWEN3_VL_8B_THINKING,
|
||||
MODEL_ID_QWEN3_VL_8B_THINKING_FP8,
|
||||
MODEL_ID_GPT_OSS_20B,
|
||||
}
|
||||
|
||||
ALLOWED_MODEL_IDS = ALLOWED_EMBEDDING_MODELS | ALLOWED_GENERATION_MODELS
|
||||
|
||||
# Default selected model (must be one of ALLOWED_MODEL_IDS).
|
||||
# If env not allowed, fallback to MODEL_ID_NOMIC.
|
||||
ENV_DEFAULT_MODEL = os.environ.get("HF_MODEL_ID", MODEL_ID_NOMIC)
|
||||
DEFAULT_MODEL_ID = (
|
||||
ENV_DEFAULT_MODEL
|
||||
if ENV_DEFAULT_MODEL in ALLOWED_MODEL_IDS
|
||||
else MODEL_ID_NOMIC
|
||||
)
|
||||
|
||||
HF_MODEL_URL = os.environ.get("HF_MODEL_URL") # optional informational field
|
||||
API_PORT = int(
|
||||
os.environ.get("PYTORCH_CONTAINER_PORT", os.environ.get("PORT", "8000"))
|
||||
)
|
||||
|
||||
# Limits (env-overridable)
|
||||
MAX_TEXTS_PER_REQUEST = int(os.environ.get("TEXT_MAX_ITEMS", "32"))
|
||||
MAX_IMAGES_PER_REQUEST = int(os.environ.get("IMAGE_MAX_ITEMS", "8"))
|
||||
MAX_IMAGE_BASE64_BYTES = int(
|
||||
os.environ.get("IMAGE_MAX_BASE64_BYTES", str(25 * 1024 * 1024))
|
||||
) # 25MB per image b64 (approx)
|
||||
MAX_IMAGE_PIXELS = int(
|
||||
os.environ.get("IMAGE_MAX_PIXELS", str(30_000_000))
|
||||
) # ~30MP safety
|
||||
|
||||
# PIL safety for large images
|
||||
ImageFile.LOAD_TRUNCATED_IMAGES = True
|
||||
Image.MAX_IMAGE_PIXELS = MAX_IMAGE_PIXELS
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
# App + Global State
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
app = FastAPI(title="AI Model Service")
|
||||
|
||||
_model_lock = threading.RLock()
|
||||
|
||||
# Unified model storage
|
||||
_model: Optional[torch.nn.Module] = None
|
||||
# Can be ColQwen2_5_Processor, AutoTokenizer, or AutoProcessor
|
||||
_processor: Optional[Any] = None
|
||||
_loaded_model_id: Optional[str] = None
|
||||
_loaded_model_type: Optional[str] = None # "embedding" or "generation"
|
||||
|
||||
# For reporting
|
||||
_dtype_str: Optional[str] = None
|
||||
_device_str: str = "cuda:0"
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
# Pydantic Models (OpenAI Compatible)
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
|
||||
class ModelCard(BaseModel):
|
||||
id: str
|
||||
object: str = "model"
|
||||
created: int = int(time.time())
|
||||
owned_by: str = "system"
|
||||
|
||||
|
||||
class ModelList(BaseModel):
|
||||
object: str = "list"
|
||||
data: List[ModelCard]
|
||||
|
||||
|
||||
class ChatMessage(BaseModel):
|
||||
role: str
|
||||
content: Union[str, List[Dict[str, Any]]] # string or multimodal list
|
||||
name: Optional[str] = None
|
||||
|
||||
|
||||
class ChatCompletionRequest(BaseModel):
|
||||
model: str
|
||||
messages: List[ChatMessage]
|
||||
temperature: Optional[float] = 1.0
|
||||
top_p: Optional[float] = 1.0
|
||||
n: Optional[int] = 1
|
||||
stream: Optional[bool] = False
|
||||
stop: Optional[Union[str, List[str]]] = None
|
||||
max_tokens: Optional[int] = None
|
||||
presence_penalty: Optional[float] = 0.0
|
||||
frequency_penalty: Optional[float] = 0.0
|
||||
logit_bias: Optional[Dict[str, float]] = None
|
||||
user: Optional[str] = None
|
||||
|
||||
|
||||
class ChatChoice(BaseModel):
|
||||
index: int
|
||||
message: ChatMessage
|
||||
finish_reason: Optional[str] = None
|
||||
|
||||
|
||||
class Usage(BaseModel):
|
||||
prompt_tokens: int
|
||||
completion_tokens: int
|
||||
total_tokens: int
|
||||
|
||||
|
||||
class ChatCompletionResponse(BaseModel):
|
||||
id: str
|
||||
object: str = "chat.completion"
|
||||
created: int
|
||||
model: str
|
||||
choices: List[ChatChoice]
|
||||
usage: Optional[Usage] = None
|
||||
|
||||
|
||||
class EmbeddingRequest(BaseModel):
|
||||
# OpenAI supports various inputs
|
||||
input: Union[str, List[str], List[int], List[List[int]]]
|
||||
model: str
|
||||
encoding_format: Optional[str] = "float" # float or base64
|
||||
user: Optional[str] = None
|
||||
|
||||
|
||||
class EmbeddingObject(BaseModel):
|
||||
object: str = "embedding"
|
||||
index: int
|
||||
# OpenAI embeddings are 1D vectors, but ColQwen is multi-vector.
|
||||
# We return the raw multi-vector as the "embedding" field,
|
||||
# which implies it's a list of lists.
|
||||
embedding: Any
|
||||
|
||||
|
||||
class EmbeddingResponse(BaseModel):
|
||||
object: str = "list"
|
||||
data: List[EmbeddingObject]
|
||||
model: str
|
||||
usage: Usage
|
||||
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
# Helpers
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _torch_dtype_str(dtype: torch.dtype) -> str:
|
||||
if dtype == torch.bfloat16:
|
||||
return "bfloat16"
|
||||
if dtype == torch.float16:
|
||||
return "float16"
|
||||
return str(dtype)
|
||||
|
||||
|
||||
def _hard_requirements_check():
|
||||
if not torch.cuda.is_available():
|
||||
raise RuntimeError(
|
||||
"CUDA is not available; a CUDA-capable GPU is required."
|
||||
)
|
||||
if not is_flash_attn_2_available():
|
||||
# Warn but maybe not fail for generation models if they can fallback?
|
||||
# But previous code had it as hard requirement. Sticking to it.
|
||||
pass
|
||||
|
||||
|
||||
def _pick_dtype() -> torch.dtype:
|
||||
return torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
|
||||
|
||||
|
||||
def _unload_model_locked():
|
||||
global _model, _processor, _loaded_model_id, _loaded_model_type, _dtype_str
|
||||
# Assumes caller holds _model_lock
|
||||
_model = None
|
||||
_processor = None
|
||||
_loaded_model_id = None
|
||||
_loaded_model_type = None
|
||||
_dtype_str = None
|
||||
gc.collect()
|
||||
if torch.cuda.is_available():
|
||||
try:
|
||||
torch.cuda.empty_cache()
|
||||
torch.cuda.ipc_collect()
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
|
||||
def _load_model_locked(model_id: str):
|
||||
global _model, _processor, _loaded_model_id, _loaded_model_type
|
||||
global _dtype_str, _device_str
|
||||
|
||||
_hard_requirements_check()
|
||||
dtype = _pick_dtype()
|
||||
device_map = "cuda:0"
|
||||
attn_impl = (
|
||||
"flash_attention_2" if is_flash_attn_2_available() else "sdpa"
|
||||
)
|
||||
|
||||
if model_id in ALLOWED_EMBEDDING_MODELS:
|
||||
# Load Embedding Model
|
||||
model = ColQwen2_5.from_pretrained(
|
||||
model_id,
|
||||
torch_dtype=dtype,
|
||||
device_map=device_map,
|
||||
attn_implementation="flash_attention_2", # ColQwen mandates FA2
|
||||
).eval()
|
||||
processor = ColQwen2_5_Processor.from_pretrained(model_id)
|
||||
_loaded_model_type = "embedding"
|
||||
|
||||
elif model_id in ALLOWED_GENERATION_MODELS:
|
||||
# Load Generation Model
|
||||
# Check if it is a VL model
|
||||
if "VL" in model_id:
|
||||
# Attempt to load as VL
|
||||
# Using AutoModelForVision2Seq or AutoModelForCausalLM
|
||||
# depending on the specific model support in transformers
|
||||
try:
|
||||
from transformers import Qwen2VLForConditionalGeneration
|
||||
|
||||
model_class = Qwen2VLForConditionalGeneration
|
||||
except ImportError:
|
||||
# Fallback to AutoModel if specific class not available
|
||||
model_class = AutoModelForCausalLM
|
||||
|
||||
# Note: We use AutoModelForCausalLM for broad compatibility.
|
||||
# Qwen2-VL requires Qwen2VLForConditionalGeneration for vision.
|
||||
# We will try AutoModelForCausalLM first.
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
model_id,
|
||||
torch_dtype=dtype,
|
||||
device_map=device_map,
|
||||
attn_implementation=attn_impl,
|
||||
trust_remote_code=True, # Often needed for new architectures
|
||||
).eval()
|
||||
|
||||
# Processor/Tokenizer
|
||||
try:
|
||||
processor = AutoProcessor.from_pretrained(
|
||||
model_id, trust_remote_code=True
|
||||
)
|
||||
except Exception:
|
||||
processor = AutoTokenizer.from_pretrained(
|
||||
model_id, trust_remote_code=True
|
||||
)
|
||||
|
||||
_loaded_model_type = "generation"
|
||||
else:
|
||||
# Standard Text Model (GPT-OSS)
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
model_id,
|
||||
torch_dtype=dtype,
|
||||
device_map=device_map,
|
||||
attn_implementation=attn_impl,
|
||||
trust_remote_code=True,
|
||||
).eval()
|
||||
processor = AutoTokenizer.from_pretrained(
|
||||
model_id, trust_remote_code=True
|
||||
)
|
||||
_loaded_model_type = "generation"
|
||||
|
||||
else:
|
||||
raise ValueError(f"Unknown model type for {model_id}")
|
||||
|
||||
_model = model
|
||||
_processor = processor
|
||||
_loaded_model_id = model_id
|
||||
_dtype_str = _torch_dtype_str(dtype)
|
||||
_device_str = device_map
|
||||
|
||||
|
||||
def _ensure_model_loaded(model_id: str):
|
||||
with _model_lock:
|
||||
if (
|
||||
_model is not None
|
||||
and _processor is not None
|
||||
and _loaded_model_id == model_id
|
||||
):
|
||||
return _model, _processor, _loaded_model_type
|
||||
|
||||
_unload_model_locked()
|
||||
_load_model_locked(model_id)
|
||||
return _model, _processor, _loaded_model_type
|
||||
|
||||
|
||||
def _current_vram_info():
|
||||
if not torch.cuda.is_available():
|
||||
return {"free": None, "total": None, "used": None}
|
||||
free, total = torch.cuda.mem_get_info(0)
|
||||
used = total - free
|
||||
return {"free": free, "total": total, "used": used}
|
||||
|
||||
|
||||
def _decode_base64_image(b64_data: str):
|
||||
approx_bytes = int(len(b64_data) * 0.75)
|
||||
if approx_bytes > MAX_IMAGE_BASE64_BYTES:
|
||||
raise ValueError(
|
||||
f"Image exceeds max base64 size of {MAX_IMAGE_BASE64_BYTES} bytes"
|
||||
)
|
||||
try:
|
||||
raw = base64.b64decode(b64_data, validate=True)
|
||||
img = Image.open(io.BytesIO(raw))
|
||||
if img.mode != "RGB":
|
||||
img = img.convert("RGB")
|
||||
img.load()
|
||||
return img
|
||||
except Exception as e:
|
||||
raise ValueError(f"Unable to decode image: {e}")
|
||||
|
||||
|
||||
def _extract_embeddings(outputs) -> torch.Tensor:
|
||||
if isinstance(outputs, torch.Tensor):
|
||||
embeddings = outputs
|
||||
elif hasattr(outputs, "last_hidden_state"):
|
||||
embeddings = outputs.last_hidden_state
|
||||
else:
|
||||
raise RuntimeError(f"Unexpected model output type: {type(outputs)}")
|
||||
|
||||
if embeddings.dim() == 2:
|
||||
embeddings = embeddings.unsqueeze(0)
|
||||
elif embeddings.dim() != 3:
|
||||
raise RuntimeError(
|
||||
f"Unexpected embedding shape: {tuple(embeddings.shape)}"
|
||||
)
|
||||
return embeddings
|
||||
|
||||
|
||||
# -----------------------------------------------------------------------------
|
||||
# Endpoints
|
||||
# -----------------------------------------------------------------------------
|
||||
|
||||
|
||||
@app.get("/health")
|
||||
def health():
|
||||
cuda_ok = bool(torch.cuda.is_available())
|
||||
flash_ok = bool(is_flash_attn_2_available())
|
||||
|
||||
info = {
|
||||
"status": "ok",
|
||||
"loaded_model_id": _loaded_model_id,
|
||||
"cuda_available": cuda_ok,
|
||||
"flash_attn_2_available": flash_ok,
|
||||
"vram_bytes": _current_vram_info(),
|
||||
}
|
||||
|
||||
if not cuda_ok:
|
||||
info["status"] = "error"
|
||||
info["error"] = "CUDA is not available."
|
||||
raise HTTPException(status_code=500, detail=info)
|
||||
|
||||
return info
|
||||
|
||||
|
||||
@app.get("/v1/models", response_model=ModelList)
|
||||
def list_models():
|
||||
models = []
|
||||
for mid in ALLOWED_MODEL_IDS:
|
||||
models.append(ModelCard(id=mid))
|
||||
return ModelList(data=models)
|
||||
|
||||
|
||||
@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
|
||||
async def chat_completions(request: ChatCompletionRequest):
|
||||
model_id = request.model
|
||||
if model_id not in ALLOWED_GENERATION_MODELS:
|
||||
raise HTTPException(
|
||||
status_code=400,
|
||||
detail=f"Model {model_id} not supported or not a generation model."
|
||||
)
|
||||
|
||||
with _model_lock:
|
||||
try:
|
||||
model, processor, mtype = _ensure_model_loaded(model_id)
|
||||
except Exception as e:
|
||||
raise HTTPException(
|
||||
status_code=500, detail=f"Failed to load model: {e}"
|
||||
)
|
||||
|
||||
if mtype != "generation":
|
||||
raise HTTPException(
|
||||
status_code=500,
|
||||
detail=(f"Model loaded as {mtype} "
|
||||
"but accessed via chat completion.")
|
||||
)
|
||||
|
||||
# Prepare input
|
||||
# Naive implementation: concatenate messages.
|
||||
# Ideally apply chat template if available.
|
||||
|
||||
prompt_text = ""
|
||||
# images = []
|
||||
|
||||
# Check if we have apply_chat_template support (most modern tokenizers do)
|
||||
has_template = hasattr(processor, "apply_chat_template")
|
||||
|
||||
if has_template:
|
||||
# processor can be Tokenizer or Processor.
|
||||
# If it is a Processor (for VL), it might expect specific format.
|
||||
# We'll try passing the messages dict directly.
|
||||
try:
|
||||
# Convert Pydantic messages to dict
|
||||
msgs = [
|
||||
m.model_dump(exclude_none=True) for m in request.messages
|
||||
]
|
||||
|
||||
# Check for images in messages if VL model
|
||||
# TODO: Extract base64 images from content if present
|
||||
|
||||
text_input = processor.apply_chat_template(
|
||||
msgs, tokenize=False, add_generation_prompt=True
|
||||
)
|
||||
except Exception as e:
|
||||
# Fallback to manual concatenation
|
||||
print(f"Template application failed: {e}")
|
||||
text_input = ""
|
||||
for m in request.messages:
|
||||
content = m.content
|
||||
if isinstance(content, list):
|
||||
# Handle multimodal content list - extract text
|
||||
content = " ".join(
|
||||
[
|
||||
c.get("text", "")
|
||||
for c in content
|
||||
if c.get("type") == "text"
|
||||
]
|
||||
)
|
||||
text_input += f"<|im_start|>{m.role}\n"
|
||||
text_input += f"{content}<|im_end|>\n"
|
||||
text_input += "<|im_start|>assistant\n"
|
||||
else:
|
||||
text_input = ""
|
||||
for m in request.messages:
|
||||
content = m.content
|
||||
if isinstance(content, list):
|
||||
content = " ".join(
|
||||
[
|
||||
c.get("text", "")
|
||||
for c in content
|
||||
if c.get("type") == "text"
|
||||
]
|
||||
)
|
||||
text_input += f"{m.role}: {content}\n"
|
||||
text_input += "assistant: "
|
||||
|
||||
# Tokenize
|
||||
inputs = None
|
||||
if (
|
||||
hasattr(processor, "process_images")
|
||||
or "Processor" in processor.__class__.__name__
|
||||
):
|
||||
# It's likely a VL processor.
|
||||
inputs = processor(
|
||||
text=[text_input], return_tensors="pt", padding=True
|
||||
).to(_device_str)
|
||||
else:
|
||||
# Standard tokenizer
|
||||
inputs = processor(text_input, return_tensors="pt").to(_device_str)
|
||||
|
||||
# Generate
|
||||
with torch.inference_mode():
|
||||
generated_ids = model.generate(
|
||||
**inputs,
|
||||
max_new_tokens=request.max_tokens or 512,
|
||||
do_sample=request.temperature > 0,
|
||||
temperature=request.temperature,
|
||||
top_p=request.top_p,
|
||||
)
|
||||
|
||||
# Decode
|
||||
input_len = inputs.input_ids.shape[1]
|
||||
generated_ids = generated_ids[:, input_len:]
|
||||
output_text = processor.decode(
|
||||
generated_ids[0], skip_special_tokens=True
|
||||
)
|
||||
|
||||
# Usage
|
||||
usage = Usage(
|
||||
prompt_tokens=input_len,
|
||||
completion_tokens=generated_ids.shape[1],
|
||||
total_tokens=input_len + generated_ids.shape[1],
|
||||
)
|
||||
|
||||
choice = ChatChoice(
|
||||
index=0,
|
||||
message=ChatMessage(role="assistant", content=output_text),
|
||||
finish_reason="stop",
|
||||
)
|
||||
|
||||
return ChatCompletionResponse(
|
||||
id=str(uuid.uuid4()),
|
||||
created=int(time.time()),
|
||||
model=model_id,
|
||||
choices=[choice],
|
||||
usage=usage,
|
||||
)
|
||||
|
||||
|
||||
@app.post("/v1/embeddings", response_model=EmbeddingResponse)
|
||||
def create_embeddings(request: EmbeddingRequest):
|
||||
model_id = request.model
|
||||
# We check if model_id is allowed.
|
||||
if model_id not in ALLOWED_EMBEDDING_MODELS:
|
||||
raise HTTPException(
|
||||
status_code=400,
|
||||
detail=f"Model {model_id} not supported or not an embedding model."
|
||||
)
|
||||
|
||||
with _model_lock:
|
||||
try:
|
||||
model, processor, mtype = _ensure_model_loaded(model_id)
|
||||
except Exception as e:
|
||||
raise HTTPException(
|
||||
status_code=500, detail=f"Failed to load model: {e}"
|
||||
)
|
||||
|
||||
if mtype != "embedding":
|
||||
raise HTTPException(
|
||||
status_code=500, detail="Model is not an embedding model."
|
||||
)
|
||||
|
||||
# Handle input
|
||||
texts = request.input
|
||||
if isinstance(texts, str):
|
||||
texts = [texts]
|
||||
# If it's list of tokens (int), we can't handle with current processor
|
||||
if texts and isinstance(texts[0], int):
|
||||
raise HTTPException(
|
||||
status_code=400,
|
||||
detail="Token IDs input not supported, please provide text.",
|
||||
)
|
||||
|
||||
try:
|
||||
with torch.inference_mode():
|
||||
# ColQwen processor handles queries/docs. Assume queries.
|
||||
batch = processor.process_queries(texts).to(_device_str)
|
||||
outputs = model(**batch)
|
||||
embeddings = _extract_embeddings(outputs)
|
||||
except Exception as exc:
|
||||
raise HTTPException(
|
||||
status_code=500, detail=f"Failed to compute embeddings: {exc}"
|
||||
)
|
||||
|
||||
embeddings = embeddings.detach().cpu().float().tolist()
|
||||
|
||||
data = []
|
||||
token_count = 0 # Dummy count
|
||||
for i, emb in enumerate(embeddings):
|
||||
data.append(EmbeddingObject(index=i, embedding=emb))
|
||||
|
||||
return EmbeddingResponse(
|
||||
data=data,
|
||||
model=model_id,
|
||||
usage=Usage(
|
||||
prompt_tokens=token_count,
|
||||
completion_tokens=0,
|
||||
total_tokens=token_count,
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import uvicorn
|
||||
|
||||
uvicorn.run(app, host="0.0.0.0", port=API_PORT)
|
||||
Reference in New Issue
Block a user