[openai api] log exception in exception handler (1/N) (#31164)

Signed-off-by: Andy Xie <andy.xning@gmail.com>
This commit is contained in:
Ning Xie
2026-03-05 16:00:12 +00:00
committed by GitHub
parent 612e7729c2
commit 176c799f4c
37 changed files with 908 additions and 1187 deletions
+201 -218
View File
@@ -54,13 +54,11 @@ class OpenAIServingEmbedding(OpenAIServing):
chat_template: str | None,
chat_template_content_format: ChatTemplateContentFormatOption,
trust_request_chat_template: bool = False,
log_error_stack: bool = False,
) -> None:
super().__init__(
engine_client=engine_client,
models=models,
request_logger=request_logger,
log_error_stack=log_error_stack,
)
self.chat_template = chat_template
@@ -75,38 +73,34 @@ class OpenAIServingEmbedding(OpenAIServing):
self,
ctx: EmbeddingServeContext,
) -> ErrorResponse | None:
try:
ctx.lora_request = self._maybe_get_adapters(ctx.request)
ctx.lora_request = self._maybe_get_adapters(ctx.request)
if isinstance(ctx.request, EmbeddingChatRequest):
error_check_ret = self._validate_chat_template(
request_chat_template=ctx.request.chat_template,
chat_template_kwargs=ctx.request.chat_template_kwargs,
trust_request_chat_template=self.trust_request_chat_template,
)
if error_check_ret is not None:
return error_check_ret
if isinstance(ctx.request, EmbeddingChatRequest):
error_check_ret = self._validate_chat_template(
request_chat_template=ctx.request.chat_template,
chat_template_kwargs=ctx.request.chat_template_kwargs,
trust_request_chat_template=self.trust_request_chat_template,
)
if error_check_ret is not None:
return error_check_ret
_, ctx.engine_prompts = await self._preprocess_chat(
ctx.request,
ctx.request.messages,
default_template=self.chat_template,
default_template_content_format=self.chat_template_content_format,
default_template_kwargs=None,
)
elif isinstance(ctx.request, EmbeddingCompletionRequest):
ctx.engine_prompts = await self._preprocess_completion(
ctx.request,
prompt_input=ctx.request.input,
prompt_embeds=None,
)
else:
return self.create_error_response("Invalid classification request type")
_, ctx.engine_prompts = await self._preprocess_chat(
ctx.request,
ctx.request.messages,
default_template=self.chat_template,
default_template_content_format=self.chat_template_content_format,
default_template_kwargs=None,
)
elif isinstance(ctx.request, EmbeddingCompletionRequest):
ctx.engine_prompts = await self._preprocess_completion(
ctx.request,
prompt_input=ctx.request.input,
prompt_embeds=None,
)
else:
return self.create_error_response("Invalid classification request type")
return None
except (ValueError, TypeError) as e:
logger.exception("Error in preprocessing prompt inputs")
return self.create_error_response(str(e))
return None
def request_output_to_embed_json_response(
self,
@@ -397,51 +391,47 @@ class OpenAIServingEmbedding(OpenAIServing):
# Custom logic for chunked processing
generators: list[AsyncGenerator[PoolingRequestOutput, None]] = []
try:
trace_headers = (
None
if ctx.raw_request is None
else await self._get_trace_headers(ctx.raw_request.headers)
trace_headers = (
None
if ctx.raw_request is None
else await self._get_trace_headers(ctx.raw_request.headers)
)
pooling_params = self._create_pooling_params(ctx)
if isinstance(pooling_params, ErrorResponse):
return pooling_params
if ctx.engine_prompts is None:
return self.create_error_response("Engine prompts not available")
max_pos_embeddings = self._get_max_position_embeddings()
for i, engine_prompt in enumerate(ctx.engine_prompts):
# Check if this specific prompt needs chunked processing
if "prompt_token_ids" in engine_prompt:
prompt_token_ids = engine_prompt["prompt_token_ids"] # type: ignore[typeddict-item]
if len(prompt_token_ids) > max_pos_embeddings:
# Use chunked processing for this prompt
chunk_generators = await self._process_chunked_request(
ctx,
prompt_token_ids,
pooling_params,
trace_headers,
i,
)
generators.extend(chunk_generators)
continue
# Normal processing for short prompts or non-token prompts
generator = await self._create_single_prompt_generator(
ctx, engine_prompt, pooling_params, trace_headers, i
)
generators.append(generator)
pooling_params = self._create_pooling_params(ctx)
if isinstance(pooling_params, ErrorResponse):
return pooling_params
ctx.result_generator = merge_async_iterators(*generators)
if ctx.engine_prompts is None:
return self.create_error_response("Engine prompts not available")
max_pos_embeddings = self._get_max_position_embeddings()
for i, engine_prompt in enumerate(ctx.engine_prompts):
# Check if this specific prompt needs chunked processing
if "prompt_token_ids" in engine_prompt:
prompt_token_ids = engine_prompt["prompt_token_ids"] # type: ignore[typeddict-item]
if len(prompt_token_ids) > max_pos_embeddings:
# Use chunked processing for this prompt
chunk_generators = await self._process_chunked_request(
ctx,
prompt_token_ids,
pooling_params,
trace_headers,
i,
)
generators.extend(chunk_generators)
continue
# Normal processing for short prompts or non-token prompts
generator = await self._create_single_prompt_generator(
ctx, engine_prompt, pooling_params, trace_headers, i
)
generators.append(generator)
ctx.result_generator = merge_async_iterators(*generators)
return None
except Exception as e:
return self.create_error_response(e)
return None
async def _collect_batch(
self,
@@ -454,164 +444,157 @@ class OpenAIServingEmbedding(OpenAIServing):
minimize memory usage.
For regular requests, collects results normally.
"""
try:
if ctx.engine_prompts is None:
return self.create_error_response("Engine prompts not available")
if ctx.engine_prompts is None:
return self.create_error_response("Engine prompts not available")
# Check if we used chunked processing
use_chunked = self._should_use_chunked_processing(ctx.request)
# Check if we used chunked processing
use_chunked = self._should_use_chunked_processing(ctx.request)
if not use_chunked:
return await super()._collect_batch(ctx=ctx)
if not use_chunked:
return await super()._collect_batch(ctx=ctx)
if ctx.result_generator is None:
return self.create_error_response("Result generator not available")
if ctx.result_generator is None:
return self.create_error_response("Result generator not available")
# Online aggregation for chunked requests to
# minimize memory usage
# Track aggregation state for each prompt
prompt_aggregators: dict[int, dict[str, Any]] = {}
short_prompts_results: dict[int, PoolingRequestOutput] = {}
# Online aggregation for chunked requests to
# minimize memory usage
# Track aggregation state for each prompt
prompt_aggregators: dict[int, dict[str, Any]] = {}
short_prompts_results: dict[int, PoolingRequestOutput] = {}
async for result_idx, result in ctx.result_generator:
if "-chunk-" in result.request_id:
# Extract prompt_idx from chunked request_id
parts = result.request_id.split("-")
try:
prompt_idx = int(parts[parts.index("prompt") + 1])
except (ValueError, IndexError):
# Fallback: extract from result_idx if parsing fails
prompt_idx = result_idx
async for result_idx, result in ctx.result_generator:
if "-chunk-" in result.request_id:
# Extract prompt_idx from chunked request_id
parts = result.request_id.split("-")
try:
prompt_idx = int(parts[parts.index("prompt") + 1])
except (ValueError, IndexError):
# Fallback: extract from result_idx if parsing fails
prompt_idx = result_idx
# Initialize aggregator for this prompt if needed
if prompt_idx not in prompt_aggregators:
prompt_aggregators[prompt_idx] = {
"weighted_sum": None,
"total_weight": 0,
"chunk_count": 0,
"request_id": result.request_id.split("-chunk-")[0],
}
# Initialize aggregator for this prompt if needed
if prompt_idx not in prompt_aggregators:
prompt_aggregators[prompt_idx] = {
"weighted_sum": None,
"total_weight": 0,
"chunk_count": 0,
"request_id": result.request_id.split("-chunk-")[0],
}
aggregator = prompt_aggregators[prompt_idx]
aggregator = prompt_aggregators[prompt_idx]
# MEAN pooling with online weighted averaging
# Ensure result is PoolingRequestOutput
# for embedding processing
if not isinstance(result, PoolingRequestOutput):
return self.create_error_response(
f"Expected PoolingRequestOutput for "
f"chunked embedding, got "
f"{type(result).__name__}"
)
# Handle both PoolingOutput and
# EmbeddingOutput types
if hasattr(result.outputs, "data"):
# PoolingOutput case
embedding_data = result.outputs.data
elif hasattr(result.outputs, "embedding"):
# EmbeddingOutput case -
# convert embedding list to tensor
embedding_data = result.outputs.embedding
else:
return self.create_error_response(
f"Unsupported output type: {type(result.outputs).__name__}"
)
if not isinstance(embedding_data, torch.Tensor):
embedding_data = torch.tensor(
embedding_data, dtype=torch.float32
)
if result.prompt_token_ids is None:
return self.create_error_response(
"prompt_token_ids cannot be None for chunked processing"
)
weight = len(result.prompt_token_ids)
weighted_embedding = embedding_data.to(dtype=torch.float32) * weight
if aggregator["weighted_sum"] is None:
# First chunk
aggregator["weighted_sum"] = weighted_embedding
else:
# Accumulate
aggregator["weighted_sum"] += weighted_embedding
aggregator["total_weight"] += weight
aggregator["chunk_count"] += 1
else:
# Non-chunked result - extract prompt_idx from request_id
parts = result.request_id.split("-")
try:
# Last part should be prompt index
prompt_idx = int(parts[-1])
except (ValueError, IndexError):
prompt_idx = result_idx # Fallback to result_idx
short_prompts_results[prompt_idx] = result
# Finalize aggregated results
final_res_batch: list[PoolingRequestOutput] = []
num_prompts = len(ctx.engine_prompts)
for prompt_idx in range(num_prompts):
if prompt_idx in prompt_aggregators:
# Finalize MEAN aggregation for this chunked prompt
aggregator = prompt_aggregators[prompt_idx]
weighted_sum = aggregator["weighted_sum"]
total_weight = aggregator["total_weight"]
if (
weighted_sum is not None
and isinstance(weighted_sum, torch.Tensor)
and isinstance(total_weight, (int, float))
and total_weight > 0
):
# Compute final mean embedding
final_embedding = weighted_sum / total_weight
# Create a PoolingRequestOutput
# for the aggregated result
pooling_output_data = PoolingOutput(data=final_embedding)
# Get original prompt token IDs for this prompt
original_prompt = ctx.engine_prompts[prompt_idx]
if "prompt_token_ids" not in original_prompt:
return self.create_error_response(
f"Chunked prompt {prompt_idx} does not contain "
"token IDs"
)
original_token_ids = original_prompt["prompt_token_ids"] # type: ignore[typeddict-item]
pooling_request_output = PoolingRequestOutput(
request_id=aggregator["request_id"],
prompt_token_ids=original_token_ids,
outputs=pooling_output_data,
num_cached_tokens=0,
finished=True,
)
final_res_batch.append(pooling_request_output)
else:
return self.create_error_response(
f"Failed to aggregate chunks for prompt {prompt_idx}"
)
elif prompt_idx in short_prompts_results:
final_res_batch.append(short_prompts_results[prompt_idx])
else:
# MEAN pooling with online weighted averaging
# Ensure result is PoolingRequestOutput
# for embedding processing
if not isinstance(result, PoolingRequestOutput):
return self.create_error_response(
f"Result not found for prompt {prompt_idx}"
f"Expected PoolingRequestOutput for "
f"chunked embedding, got "
f"{type(result).__name__}"
)
ctx.final_res_batch = final_res_batch
# Handle both PoolingOutput and
# EmbeddingOutput types
if hasattr(result.outputs, "data"):
# PoolingOutput case
embedding_data = result.outputs.data
elif hasattr(result.outputs, "embedding"):
# EmbeddingOutput case -
# convert embedding list to tensor
embedding_data = result.outputs.embedding
else:
return self.create_error_response(
f"Unsupported output type: {type(result.outputs).__name__}"
)
return None
if not isinstance(embedding_data, torch.Tensor):
embedding_data = torch.tensor(embedding_data, dtype=torch.float32)
except Exception as e:
return self.create_error_response(e)
if result.prompt_token_ids is None:
return self.create_error_response(
"prompt_token_ids cannot be None for chunked processing"
)
weight = len(result.prompt_token_ids)
weighted_embedding = embedding_data.to(dtype=torch.float32) * weight
if aggregator["weighted_sum"] is None:
# First chunk
aggregator["weighted_sum"] = weighted_embedding
else:
# Accumulate
aggregator["weighted_sum"] += weighted_embedding
aggregator["total_weight"] += weight
aggregator["chunk_count"] += 1
else:
# Non-chunked result - extract prompt_idx from request_id
parts = result.request_id.split("-")
try:
# Last part should be prompt index
prompt_idx = int(parts[-1])
except (ValueError, IndexError):
prompt_idx = result_idx # Fallback to result_idx
short_prompts_results[prompt_idx] = result
# Finalize aggregated results
final_res_batch: list[PoolingRequestOutput] = []
num_prompts = len(ctx.engine_prompts)
for prompt_idx in range(num_prompts):
if prompt_idx in prompt_aggregators:
# Finalize MEAN aggregation for this chunked prompt
aggregator = prompt_aggregators[prompt_idx]
weighted_sum = aggregator["weighted_sum"]
total_weight = aggregator["total_weight"]
if (
weighted_sum is not None
and isinstance(weighted_sum, torch.Tensor)
and isinstance(total_weight, (int, float))
and total_weight > 0
):
# Compute final mean embedding
final_embedding = weighted_sum / total_weight
# Create a PoolingRequestOutput
# for the aggregated result
pooling_output_data = PoolingOutput(data=final_embedding)
# Get original prompt token IDs for this prompt
original_prompt = ctx.engine_prompts[prompt_idx]
if "prompt_token_ids" not in original_prompt:
return self.create_error_response(
f"Chunked prompt {prompt_idx} does not contain token IDs"
)
original_token_ids = original_prompt["prompt_token_ids"] # type: ignore[typeddict-item]
pooling_request_output = PoolingRequestOutput(
request_id=aggregator["request_id"],
prompt_token_ids=original_token_ids,
outputs=pooling_output_data,
num_cached_tokens=0,
finished=True,
)
final_res_batch.append(pooling_request_output)
else:
return self.create_error_response(
f"Failed to aggregate chunks for prompt {prompt_idx}"
)
elif prompt_idx in short_prompts_results:
final_res_batch.append(short_prompts_results[prompt_idx])
else:
return self.create_error_response(
f"Result not found for prompt {prompt_idx}"
)
ctx.final_res_batch = final_res_batch
return None
async def create_embedding(
self,