Merge pull request #203 from InfiniTensor/issue/193

Issue/193: inference_server适配部署需求

python/infinilm/llm/cache_manager.py

@@ -261,6 +261,12 @@ class BlockManager:
     def get_num_free_blocks(self) -> int:
         return len(self.free_block_ids)
 
+    def get_total_usable_blocks(self) -> int:
+        freeable_used_blocks = sum(
+            1 for bid in self.used_block_ids if self.blocks[bid].ref_count == 0
+        )
+        return len(self.free_block_ids) + freeable_used_blocks
+
     def __repr__(self):
         return (
             f"BlockManager(blocks={self.num_blocks}, block_size={self.block_size}, "

python/infinilm/llm/llm.py

@@ -228,12 +228,63 @@ class LLMEngine:
             req.generated_token_ids.append(token_id)
             if req.is_prefill:
                 req.is_prefill = False
-
-            token_text = self.tokenizer.decode(token_id)
+            # vLLM-style replacement character handling is primarily relevant for streaming.
+            # For offline generation (no output queue), keep the fast incremental path.
+            if req._output_queue is None:
+                token_text = self.detokenize([token_id])
                 req.generated_text += token_text
+            else:
+                # Streaming path: compute delta from a full decode so we can hold back
+                # trailing '\ufffd' (likely an incomplete UTF-8 sequence).
+                decoded_text = self.detokenize(req.generated_token_ids)
+
+                finished_now = False
+                # Update generated_text to the latest decode (used for stop-string checks and debugging)
+                req.generated_text = decoded_text
 
                 if self._check_request_finished(req, token_id):
                     req.mark_finished(req.finish_reason)
+                    finished_now = True
+
+                    # Remove stop string from generated_text if STOP_STRING finish reason
+                    if req.finish_reason == FinishReason.STOP_STRING:
+                        stop_strings = req.sampling_params.stop or []
+                        for stop_str in stop_strings:
+                            if decoded_text.endswith(stop_str):
+                                # Remove the stop string from the end
+                                decoded_text = decoded_text[:-len(stop_str)]
+                                req.generated_text = decoded_text
+                                break
+
+                holds_back_incomplete_utf8 = (
+                    bool(decoded_text) and decoded_text.endswith("\ufffd")
+                )
+
+                # vLLM-style: hold back only if we are not on the final chunk.
+                # Suppress output when finish reason is LENGTH or STOP_STRING.
+                # Root cause fix: When STOP_STRING is detected, we suppress output for the token
+                # that completes the stop string, preventing additional tokens from being output.
+                if (holds_back_incomplete_utf8 and not finished_now) or (
+                    finished_now and req.finish_reason in (FinishReason.LENGTH, FinishReason.STOP_STRING)
+                ):
+                    token_text = ""
+                else:
+                    last_len = getattr(req, "_stream_last_yielded_length", 0)
+                    token_text = decoded_text[last_len:]
+                    if token_text:
+                        req._stream_last_yielded_length = len(decoded_text)
+
+            # For non-streaming, finish checks happen here.
+            if req._output_queue is None and self._check_request_finished(req, token_id):
+                req.mark_finished(req.finish_reason)
+                # Remove stop string from generated_text if STOP_STRING finish reason
+                if req.finish_reason == FinishReason.STOP_STRING:
+                    stop_strings = req.sampling_params.stop or []
+                    for stop_str in stop_strings:
+                        if req.generated_text.endswith(stop_str):
+                            # Remove the stop string from the end
+                            req.generated_text = req.generated_text[:-len(stop_str)]
+                            break
 
             # Put output in queue if it exists (for async streaming)
             if req._output_queue is not None:
@@ -283,12 +334,15 @@ class LLMEngine:
         self,
         messages: List[dict],
         add_generation_prompt: bool = True,
+        chat_template_kwargs: Optional[dict] = None,
     ) -> str:
         """Apply chat template to messages."""
+        chat_template_kwargs = chat_template_kwargs or {}
         return self.tokenizer.apply_chat_template(
             conversation=messages,
             add_generation_prompt=add_generation_prompt,
             tokenize=False,
+            **chat_template_kwargs,
         )
 
 
@@ -486,6 +540,10 @@ class AsyncLLMEngine:
 
         self._running = False
         self._step_thread: Optional[threading.Thread] = None
+        self._healthy = True
+
+    def is_healthy(self) -> bool:
+        return bool(self._healthy)
 
     def start(self):
         """Start the background inference loop."""
@@ -520,6 +578,7 @@ class AsyncLLMEngine:
                     time.sleep(0.01)
             except Exception as e:
                 logger.error(f"Error in step loop: {e}", exc_info=True)
+                self._healthy = False
                 self._running = False
                 break
 
@@ -581,6 +640,8 @@ class AsyncLLMEngine:
         request_id: Optional[str] = None,
         request_data: Optional[dict] = None,
         http_request: Optional[any] = None,
+        add_generation_prompt: bool = True,
+        chat_template_kwargs: Optional[dict] = None,
     ) -> InferenceRequest:
         """Add a chat request to the engine.
 
@@ -594,7 +655,11 @@ class AsyncLLMEngine:
         Returns:
             The created InferenceRequest object.
         """
-        prompt = self.engine.apply_chat_template(messages, add_generation_prompt=True)
+        prompt = self.engine.apply_chat_template(
+            messages,
+            add_generation_prompt=add_generation_prompt,
+            chat_template_kwargs=chat_template_kwargs,
+        )
         return self.add_request(
             prompt=prompt,
             sampling_params=sampling_params,
@@ -607,6 +672,7 @@ class AsyncLLMEngine:
         self,
         request: InferenceRequest,
         timeout: float = 100.0,
+        request_timeout: Optional[float] = None,
     ) -> AsyncIterator[TokenOutput]:
         """Stream tokens from a request.
 
@@ -619,6 +685,7 @@ class AsyncLLMEngine:
         """
         import asyncio
 
+        start = time.time()
         while True:
             if request.is_finished() and request.output_queue.async_q.empty():
                 break
@@ -635,6 +702,20 @@ class AsyncLLMEngine:
                 if token_output.finished:
                     break
             except asyncio.TimeoutError:
+                # Enforce request-level timeout even if no tokens are produced.
+                if request_timeout is not None:
+                    now = time.time()
+                    if now - start > float(request_timeout):
+                        request.mark_timeout()
+                        yield TokenOutput(
+                            request_id=request.request_id,
+                            token_id=-1,
+                            token_text="",
+                            finished=True,
+                            finish_reason=FinishReason.TIMEOUT,
+                            generated_text=request.generated_text,
+                        )
+                        break
                 if request.is_finished():
                     break
                 continue

python/infinilm/llm/request.py

@@ -144,6 +144,10 @@ class InferenceRequest:
         # Output management (for async streaming)
         self._output_queue: Optional[janus.Queue] = None
 
+        # Streaming helpers (vLLM-style UTF-8 buffering at the chunking layer)
+        # Used by the engine to compute "delta" text chunks from a full decode.
+        self._stream_last_yielded_length: int = 0
+
     @property
     def output_queue(self) -> janus.Queue:
         """Lazy initialization of output queue."""

python/infinilm/llm/sampling_params.py

@@ -15,7 +15,7 @@ class SamplingParams:
     top_k: int = 1
     max_tokens: Optional[int] = None
     stop: Optional[List[str]] = None
-    stop_token_ids: Optional[List[int]] = None
+    stop_token_ids: Optional[List[int]] = None  # Placeholder for future usage, not currently handled
 
     def __post_init__(self):
         if self.stop is None:

python/infinilm/llm/scheduler.py

@@ -155,12 +155,21 @@ class Scheduler:
             except queue.Empty:
                 break
 
+            if not self.can_accept_request(req):
+                self.waiting_queue.sync_q.put(req)
+                break
+
+            # Skip requests that were already finished (e.g., timed out/canceled while waiting)
+            if req.is_finished():
+                self.complete_requests([req])
+                continue
+
             req_tokens = req.get_input_tokens()
             num_required_blocks = req.get_num_blocks_required(self.block_size)
 
             if not self.cache_manager.can_allocate(num_required_blocks):
                 if not self.cache_manager.try_free_blocks(num_required_blocks):
-                    raise RuntimeError("No available cache blocks")
+                    raise RuntimeError("No available cache blocks for new request")
 
             # Allocate blocks with automatic prefix caching support
             req.block_table, req.slot_mapping, req.num_cached_tokens = (
@@ -185,6 +194,10 @@ class Scheduler:
                 req = self.running_queue.sync_q.get_nowait()
             except queue.Empty:
                 break
+            # Skip requests that were already finished (e.g., timed out/canceled while running)
+            if req.is_finished():
+                self.complete_requests([req])
+                continue
 
             # Decode phase: allocate slot for newly generated token
             try:
@@ -197,7 +210,7 @@ class Scheduler:
                 scheduled_requests.append(req)
 
             except RuntimeError as e:
-                raise RuntimeError("No available cache blocks") from e
+                raise RuntimeError("No available cache blocks for new token") from e
 
         # Return decode batch if any running requests were scheduled
         if scheduled_requests:
@@ -237,6 +250,31 @@ class Scheduler:
                 # Still running, put back in running queue
                 self.running_queue.sync_q.put(req)
 
+    def can_accept_request(self, request: InferenceRequest) -> bool:
+        total_required_blocks = 0
+
+        # Calculate blocks needed for running requests
+        running_queue_size = self.running_queue.sync_q.qsize()
+        for _ in range(running_queue_size):
+            req = self.running_queue.sync_q.get()
+            remaining_tokens = (
+                req.sampling_params.max_tokens - req.get_num_generated_tokens()
+            )
+            num_blocks_needed = (
+                remaining_tokens + self.block_size - 1
+            ) // self.block_size
+            total_required_blocks += num_blocks_needed
+            self.running_queue.sync_q.put(req)
+
+        # Calculate blocks needed for the new request
+        total_length = request.get_prompt_length()
+        total_length += request.sampling_params.max_tokens
+        num_blocks_needed = (total_length + self.block_size - 1) // self.block_size
+        total_required_blocks += num_blocks_needed
+
+        # Compare with total usable blocks in cache manager
+        return total_required_blocks <= self.cache_manager.get_total_usable_blocks()
+
     def get_cache_stats(self) -> dict:
         """Get cache statistics."""
         return {

python/infinilm/server/inference_server.py

@@ -10,6 +10,7 @@ import uuid
 import argparse
 import uvicorn
 import logging
+import os
 
 from fastapi import FastAPI, Request
 from fastapi.responses import JSONResponse, StreamingResponse
@@ -22,7 +23,7 @@ DEFAULT_STREAM_TIMEOUT = 100.0
 DEFAULT_REQUEST_TIMEOUT = 1000.0
 
 
-def chunk_json(id_, content=None, role=None, finish_reason=None):
+def chunk_json(id_, content=None, role=None, finish_reason=None, model: str = "unknown"):
     """Generate JSON chunk for streaming response."""
     delta = {}
     if content:
@@ -33,7 +34,7 @@ def chunk_json(id_, content=None, role=None, finish_reason=None):
         "id": id_,
         "object": "chat.completion.chunk",
         "created": int(time.time()),
-        "model": "jiuge",
+        "model": model,
         "system_fingerprint": None,
         "choices": [
             {
@@ -84,6 +85,8 @@ class InferenceServer:
             port: Server port number.
         """
         self.model_path = model_path
+        # vLLM-like served model id: directory name of model_path
+        self.model_id = os.path.basename(os.path.normpath(model_path)) or "model"
         self.device = device
         self.dtype = dtype
         self.tensor_parallel_size = tensor_parallel_size
@@ -136,7 +139,10 @@ class InferenceServer:
     def _register_routes(self, app: FastAPI):
         """Register API routes."""
 
+        # OpenAI-compatible chat completions endpoint.
+        # Support both legacy path and OpenAI-style /v1 prefix for proxy/router compatibility.
         @app.post("/chat/completions")
+        @app.post("/v1/chat/completions")
         async def chat_completions(request: Request):
             try:
                 data = await request.json()
@@ -169,15 +175,21 @@ class InferenceServer:
 
         @app.get("/health")
         async def health():
+            # Expose engine health so babysitter/registry can treat backend as unhealthy.
+            if (
+                self.engine is not None
+                and hasattr(self.engine, "is_healthy")
+                and not self.engine.is_healthy()
+            ):
+                return JSONResponse(content={"status": "unhealthy"}, status_code=503)
             return {"status": "healthy"}
 
-        @app.get("/v1/models")
-        async def list_models():
+        def _models_payload():
             return {
                 "object": "list",
                 "data": [
                     {
-                        "id": "jiuge",
+                        "id": self.model_id,
                         "object": "model",
                         "created": int(time.time()),
                         "owned_by": "infinilm",
@@ -185,20 +197,54 @@ class InferenceServer:
                 ],
             }
 
+        # Support both /v1/models (OpenAI) and /models (common legacy) for compatibility.
+        @app.get("/v1/models")
+        async def list_models():
+            return _models_payload()
+
+        @app.get("/models")
+        async def list_models_legacy():
+            return _models_payload()
+
     def _build_sampling_params(self, data: dict) -> SamplingParams:
         """Build SamplingParams from request data."""
+        # Support both:
+        # - top-level OpenAI-ish fields: temperature/top_p/top_k/max_tokens/stop
+        # - nested dict: sampling_params: { ... }
+        sp = data.get("sampling_params") or {}
+        if not isinstance(sp, dict):
+            sp = {}
+
+        def pick(key: str, default):
+            # Priority: explicit top-level field > nested sampling_params > server default
+            if key in data and data.get(key) is not None:
+                return data.get(key)
+            if key in sp and sp.get(key) is not None:
+                return sp.get(key)
+            return default
+
+        # Accept common alias
+        max_tokens = pick("max_tokens", self.max_tokens)
+        if max_tokens is None:
+            # Some clients use max_new_tokens
+            max_tokens = pick("max_new_tokens", self.max_tokens)
+
+        stop = pick("stop", None)
+        if isinstance(stop, str):
+            stop = [stop]
+
+
         return SamplingParams(
-            temperature=data.get("temperature", self.temperature),
-            top_p=data.get("top_p", self.top_p),
-            top_k=data.get("top_k", self.top_k),
-            max_tokens=data.get("max_tokens", self.max_tokens),
-            stop=data.get("stop"),
+            temperature=float(pick("temperature", self.temperature)),
+            top_p=float(pick("top_p", self.top_p)),
+            top_k=int(pick("top_k", self.top_k)),
+            max_tokens=int(max_tokens) if max_tokens is not None else None,
+            stop=stop,
         )
 
     async def _stream_chat(self, request_id: str, data: dict, http_request: Request):
         """Handle streaming chat request."""
         req = None
-        start_time = time.time()
 
         try:
             messages = data.get("messages", [])
@@ -210,22 +256,26 @@ class InferenceServer:
                 request_id=request_id,
                 request_data=data,
                 http_request=http_request,
+                add_generation_prompt=bool(data.get("add_generation_prompt", True)),
+                chat_template_kwargs=data.get("chat_template_kwargs") or {},
             )
 
             async for token_output in self.engine.stream_request(
-                req, timeout=DEFAULT_STREAM_TIMEOUT
+                req,
+                timeout=DEFAULT_STREAM_TIMEOUT,
+                request_timeout=DEFAULT_REQUEST_TIMEOUT,
             ):
-                # Check timeout
-                if time.time() - start_time > DEFAULT_REQUEST_TIMEOUT:
+                # If stream_request enforces timeout, we can just surface the state to the client.
+                if token_output.finish_reason == FinishReason.TIMEOUT:
                     logger.warning(
                         f"Request {request_id} timed out after {DEFAULT_REQUEST_TIMEOUT}s"
                     )
-                    req.mark_timeout()
                     error_chunk = json.dumps(
                         chunk_json(
                             request_id,
                             content="[Request timeout]",
                             finish_reason="timeout",
+                            model=self.model_id,
                         ),
                         ensure_ascii=False,
                     )
@@ -238,9 +288,19 @@ class InferenceServer:
                     req.mark_canceled()
                     break
 
+                # Skip EOS token text for OpenAI API compatibility
+                # Check if this token is an EOS token by comparing token_id with eos_token_ids
+                eos_token_ids = self.engine.engine.eos_token_ids
+                is_eos_token = (
+                    eos_token_ids and token_output.token_id in eos_token_ids
+                )
+
+                if not is_eos_token and token_output.token_text:
                     # Send token
                     chunk = json.dumps(
-                    chunk_json(request_id, content=token_output.token_text),
+                        chunk_json(
+                            request_id, content=token_output.token_text, model=self.model_id
+                        ),
                         ensure_ascii=False,
                     )
                     yield f"data: {chunk}\n\n"
@@ -250,7 +310,9 @@ class InferenceServer:
                         token_output.finish_reason
                     )
                     chunk = json.dumps(
-                        chunk_json(request_id, finish_reason=finish_reason),
+                        chunk_json(
+                            request_id, finish_reason=finish_reason, model=self.model_id
+                        ),
                         ensure_ascii=False,
                     )
                     yield f"data: {chunk}\n\n"
@@ -262,7 +324,10 @@ class InferenceServer:
                 req.mark_failed()
             error_chunk = json.dumps(
                 chunk_json(
-                    request_id, content=f"[Error: {str(e)}]", finish_reason="error"
+                    request_id,
+                    content=f"[Error: {str(e)}]",
+                    finish_reason="error",
+                    model=self.model_id,
                 ),
                 ensure_ascii=False,
             )
@@ -278,7 +343,6 @@ class InferenceServer:
     async def _chat(self, request_id: str, data: dict, http_request: Request):
         """Handle non-streaming chat request."""
         req = None
-        start_time = time.time()
 
         try:
             messages = data.get("messages", [])
@@ -290,17 +354,20 @@ class InferenceServer:
                 request_id=request_id,
                 request_data=data,
                 http_request=http_request,
+                add_generation_prompt=bool(data.get("add_generation_prompt", True)),
+                chat_template_kwargs=data.get("chat_template_kwargs") or {},
             )
 
             # Collect all generated tokens
             output_text = ""
             async for token_output in self.engine.stream_request(
-                req, timeout=DEFAULT_STREAM_TIMEOUT
+                req,
+                timeout=DEFAULT_STREAM_TIMEOUT,
+                request_timeout=DEFAULT_REQUEST_TIMEOUT,
             ):
-                # Check timeout
-                if time.time() - start_time > DEFAULT_REQUEST_TIMEOUT:
+                # Request-level timeout is handled inside stream_request.
+                if token_output.finish_reason == FinishReason.TIMEOUT:
                     logger.warning(f"Request {request_id} timed out")
-                    req.mark_timeout()
                     break
 
                 # Check client disconnect
@@ -309,6 +376,14 @@ class InferenceServer:
                     req.mark_canceled()
                     break
 
+                # Skip EOS token text for OpenAI API compatibility
+                # Check if this token is an EOS token by comparing token_id with eos_token_ids
+                eos_token_ids = self.engine.engine.eos_token_ids
+                is_eos_token = (
+                    eos_token_ids and token_output.token_id in eos_token_ids
+                )
+
+                if not is_eos_token:
                     output_text += token_output.token_text
 
                 if token_output.finished:
@@ -322,6 +397,7 @@ class InferenceServer:
                 content=output_text,
                 role="assistant",
                 finish_reason=finish_reason or "stop",
+                model=self.model_id,
             )
             return response
 

test/bench/test_benchmark.py

@@ -4,7 +4,6 @@ import argparse
 import time
 import re
 import csv
-from datasets import load_dataset, Dataset
 import numpy as np
 import infinicore
 from infinilm.modeling_utils import load_model_state_dict_by_file
@@ -12,6 +11,7 @@ from infinilm.distributed import DistConfig
 from infinilm.cache import StaticKVCacheConfig
 from infinilm.infer_engine import GenerationConfig, InferEngine
 from infinilm.cache import StaticKVCacheConfig
+from datasets import load_dataset, Dataset
 from abc import ABC, abstractmethod