From 8b98b7d0b465920740cfa1acf5dde1b6a57d8d52 Mon Sep 17 00:00:00 2001 From: Martin Hickey Date: Mon, 21 Sep 2026 15:13:52 +0100 Subject: [PATCH] [Frontend] Add streaming parity tests and docs for derender (#57922) Signed-off-by: Martin Hickey --- docs/serving/online_serving/derenderer.md | 111 +++++++++- .../derender/test_derender_parity.py | 189 +++++++++++++++++- .../derender/test_derender_stream.py | 104 ++-------- tests/entrypoints/scale_out/derender/utils.py | 103 ++++++++++ .../scale_out/token_in_token_out/protocol.py | 4 + 5 files changed, 410 insertions(+), 101 deletions(-) create mode 100644 tests/entrypoints/scale_out/derender/utils.py diff --git a/docs/serving/online_serving/derenderer.md b/docs/serving/online_serving/derenderer.md index 773131d033..0e865796f8 100644 --- a/docs/serving/online_serving/derenderer.md +++ b/docs/serving/online_serving/derenderer.md @@ -6,7 +6,7 @@ This closes the loop for a token-in / token-out engine in disaggregated serving: - **GPU less post processing**: Detokenization, reasoning parsing, and tool call parsing run on the same GPU less frontend that hosts `/render` - **Parser parity**: The derenderer reuses vLLM's tool and reasoning parsers, so a disaggregated deployment produces the same `content`/`reasoning`/ `tool_calls` split as a standard `vllm serve` server -- **Streaming and one-shot parsing**: Non-streaming calls send a complete `GenerateResponse` with all token IDs present and get one-shot parsing. Both endpoints also accept `stream: true`, taking one `GenerateStreamResponse` delta plus a client carried `stream_state` and returning `{chunk, stream_state}`. The chat endpoint's streaming path supports reasoning and tool call parsing while emitting the same `reasoning`/`content`/`tool_calls` deltas the generate streaming path would for the same token IDs +- **Streaming and one-shot parsing**: Non-streaming calls send a complete `GenerateResponse` with all token IDs present and get one-shot parsing. Both endpoints also accept `stream: true`, taking one `GenerateStreamResponse` delta plus a client carried `stream_state` and returning `{chunk, stream_state}`. The chat endpoint's streaming path supports reasoning and tool call parsing while emitting the same `reasoning`/`content`/`tool_calls` deltas the generate streaming path would for the same token IDs (see [Streaming](#streaming)) Both endpoints are hosted by the GPU less rendering server started with [`vllm launch render`](../../cli/launch/render.md), alongside the `/render` endpoints. @@ -27,8 +27,10 @@ The derender step needs more than the engine's `token_ids`. It also consumes the - Chat Completions Derender API (`/v1/chat/completions/derender`) - Post process a single `GenerateResponse` into a `ChatCompletionResponse` + - With `stream: true`, post process one `GenerateStreamResponse` chunk into a `ChatCompletionStreamResponse` chunk - Completions Derender API (`/v1/completions/derender`) - Post process a list of `GenerateResponse` objects (one per prompt) into a `CompletionResponse` + - With `stream: true`, post process one `GenerateStreamResponse` chunk into a `CompletionStreamResponse` chunk ## Request format @@ -50,8 +52,54 @@ Each request wraps the engine's `GenerateResponse`(s) together with the caller m --8<-- "vllm/entrypoints/scale_out/token_in_token_out/protocol.py:derender-completion-request" ``` +Streaming requests set `stream: true` and carry one generate chunk (`generate_chunk`) plus `stream_state` instead of a complete response. + +`/v1/chat/completions/derender` with `stream: true`: + +??? code + + ```python + --8<-- "vllm/entrypoints/scale_out/token_in_token_out/protocol.py:derender-chat-stream-request" + ``` + +`/v1/completions/derender` with `stream: true`: + +??? code + + ```python + --8<-- "vllm/entrypoints/scale_out/token_in_token_out/protocol.py:derender-completion-stream-request" + ``` + +Both return `{"chunk": ..., "stream_state": ...}`. `chunk` is a `ChatCompletionStreamResponse` or `CompletionStreamResponse` and `stream_state` goes with the next call. + Oversized payloads are rejected with a `400` before any `tokenizer.decode()` or parser runs. +## Parser configuration + +The derenderer builds its tool and reasoning parsers from its own server flags plus the `chat_request` sent with each call. It can't see how the request was rendered or served anywhere else, so for its output to match a `vllm serve` server: + +- Start the render server with the same `--tool-call-parser`, `--reasoning-parser`, `--enable-auto-tool-choice`, `--chat-template` and `--default-chat-template-kwargs` you'd give `vllm serve` for this model. If `/render` and `/derender` run on different servers, give both the same values. +- Send the full `chat_request` that went to `/render`, not just `messages` and `tools`. Fields like `chat_template_kwargs`, `reasoning_effort`, `tool_choice` and `include_reasoning` change how the output is parsed. + +A mismatch doesn't fail. The parser just splits `reasoning`, `content` and `tool_calls` differently from what `vllm serve` would return. + +`/v1/completions/derender` only detokenizes. It never runs tool or reasoning parsers, the same as `/v1/completions` on `vllm serve`. It only reads `skip_special_tokens` from `completion_request`. + +## Streaming + +Streaming derender mirrors `/inference/v1/generate`. Set `stream: true` in the body on the same path and the endpoint takes one generate stream chunk instead of a complete response. It answers with JSON, not SSE. Each call turns one generate chunk into one derendered chunk and the client forwards that to its own caller. + +The server keeps no state between calls. Everything the next call needs is in `stream_state` which the client carries: + +- **Echo `stream_state` back.** Leave it out (or send `null`) on the first call, then send the `stream_state` from each response with the next request, unchanged. State that doesn't add up (e.g. `output_chunk_lens` that don't sum to the number of output tokens) is rejected with a `400`. +- **One choice per chunk.** `/inference/v1/generate` sends one choice per SSE event and each derender call accepts at most one. More than one is rejected with a `400`. +- **`n > 1` is N streams.** Generate interleaves chunks for different choices and tags each with its `index`. Keep one `stream_state` per index and send each chunk with the state for its index. Each index gets its own `role` delta on its first chunk. +- **Send every chunk through, including the last ones.** The finish reason usually arrives with the final tokens. A chunk with a `finish_reason` and no tokens is accepted too and flushes any buffered tool call arguments. With `stream_options: {"include_usage": true}`, generate ends with a usage only chunk (`choices: []`). Send it through derender like any other chunk to get the usage chunk for the response. `usage.prompt_tokens` is the request's `prompt_tokens` if set, otherwise the generate chunk's. +- **Send the same context on every call.** `chat_request` and `prompt_token_ids` aren't kept between calls, so they go with every chunk including the usage chunk. Both are required when a tool or reasoning parser is configured. `prompt_token_ids` is the `token_ids` of the `GenerateRequest` returned by `/render`. +- **Don't forward `[DONE]`.** It marks the end of the generate stream and isn't a chunk. + +Streaming chunks don't carry logprobs yet, so `logprobs` on a generate chunk are dropped. The non-streaming endpoints do resolve them, including `token_id:N` placeholders. + ## Streaming cost Streaming derender threads a client carried `stream_state` across per-chunk calls instead of keeping session state on the server. @@ -118,3 +166,64 @@ print(response["choices"][0]["message"]["content"]) ``` Passing `chat_request` lets the derenderer run the configured tool and reasoning parsers. This means `response["choices"][0]["message"]` carries the same `content` / `reasoning` / `tool_calls` split a `vllm serve` server would produce. `chat_request` can only be omitted for a model with no tool or reasoning parser configured. A parser configured model rejects a missing `chat_request` with a 400 rather than silently falling back to plain detokenization. + +### Streaming example + +The same round trip, streamed. `/render` keeps `stream` and `stream_options` on the `GenerateRequest` it returns, so the generate call streams too. Each generate chunk goes through derender with the `stream_state` from the previous call (see [Streaming](#streaming)). + +```python +import json + +import httpx + +MODEL = "meta-llama/Llama-3.2-1B-Instruct" +RENDER = "http://localhost:8100" # vllm launch render ... +ENGINE = "http://localhost:8200" # token-in / token-out engine + +chat_request = { + "model": MODEL, + "messages": [{"role": "user", "content": "What is 2+2?"}], + "max_tokens": 32, + "stream": True, + "stream_options": {"include_usage": True}, +} + +with httpx.Client(timeout=60.0) as client: + # 1. Render: request -> token IDs (GPU less) + generate_request = client.post( + f"{RENDER}/v1/chat/completions/render", json=chat_request + ).json() + prompt_token_ids = generate_request["token_ids"] + + # 2. Generate: stream token IDs (token-in / token-out engine) + stream_state = None # one per choice index when n > 1 + with client.stream( + "POST", f"{ENGINE}/inference/v1/generate", json=generate_request + ) as generate_stream: + for line in generate_stream.iter_lines(): + if not line.startswith("data: ") or line == "data: [DONE]": + continue + + # 3. Derender: one generate chunk -> one chat completion chunk + derendered = client.post( + f"{RENDER}/v1/chat/completions/derender", + json={ + "stream": True, + "model": MODEL, + "generate_chunk": json.loads(line[len("data: ") :]), + "stream_state": stream_state, + "prompt_tokens": len(prompt_token_ids), + "prompt_token_ids": prompt_token_ids, + "chat_request": chat_request, + }, + ).json() + stream_state = derendered["stream_state"] + + chunk = derendered["chunk"] + for choice in chunk["choices"]: + print(choice["delta"].get("content") or "", end="", flush=True) + if chunk.get("usage"): + print(f"\n{chunk['usage']}") +``` + +With a tool or reasoning parser configured, `delta` also carries `reasoning` and `tool_calls`, the same deltas `/v1/chat/completions` streams for those tokens. diff --git a/tests/entrypoints/scale_out/derender/test_derender_parity.py b/tests/entrypoints/scale_out/derender/test_derender_parity.py index adece09c0a..a89f8d68db 100644 --- a/tests/entrypoints/scale_out/derender/test_derender_parity.py +++ b/tests/entrypoints/scale_out/derender/test_derender_parity.py @@ -3,20 +3,25 @@ """Round trip parity CI: render -> generate -> derender -> render == render. -Pins the coupled parsing path (``/v1/chat/completions`` on a normal GPU +Pins the coupled parsing path (`/v1/chat/completions` on a normal GPU server which parses generated tokens incrementally as they stream out) -against the disaggregated path (``OnlineDerenderer.derender_chat`` via -``/v1/chat/completions/derender`` which parses the same tokens all at once, -out of process). A standard ``vllm serve`` GPU server mounts both, so one +against the disaggregated path (`OnlineDerenderer.derender_chat` via +`/v1/chat/completions/derender` which parses the same tokens all at once, +out of process). A standard `vllm serve` GPU server mounts both, so one real generation lets both parsers run on identical input. Both paths consume the same generated token IDs (extracted from the coupled -response's ``token_ids`` via ``return_token_ids=True``), so generation +response's `token_ids` via `return_token_ids=True`), so generation nondeterminism is irrelevant. The only variable under test is whether the two parsing code paths agree. Parity is asserted unconditionally as only the stronger per case assertions (e.g. "a tool call was produced") are gated behind the marker actually having been emitted since a 1.5B model is not -guaranteed to emit ```` / ````. +guaranteed to emit `` / ``. + +The streaming cases do the same for `stream=true`. Each coupled chunk's +`token_ids` becomes one `/inference/v1/generate` stream chunk, fed +through the chunked derender path with `stream_state` threaded across calls +the way a client would, so both parsers see the same chunk boundaries. """ import json @@ -24,7 +29,12 @@ import json import httpx import pytest import pytest_asyncio +import regex as re +from tests.entrypoints.scale_out.derender.utils import ( + assemble_stream, + stream_chat_derender, +) from tests.utils import RemoteOpenAIServer MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" @@ -228,8 +238,8 @@ async def test_parity_logprobs(client): """token_id:N resolution parity vs. the coupled server's real strings. A real disaggregated worker only has token IDs so it emits logprobs - with ``token_id:N`` placeholders (``return_tokens_as_token_ids=True`` - reproduces that shape here). ``/derender`` must resolve those + with `token_id:N` placeholders (`return_tokens_as_token_ids=True` + reproduces that shape here). `/derender` must resolve those placeholders to the same token strings/bytes the coupled server resolves them to directly. """ @@ -267,3 +277,166 @@ async def test_parity_logprobs(client): for d_entry, r_entry in zip(d_content, r_content): assert d_entry["token"] == r_entry["token"] assert d_entry["bytes"] == r_entry["bytes"] + + +# --------------------------------------------------------------------------- +# Streaming helpers +# --------------------------------------------------------------------------- + + +async def _coupled_stream( + client: httpx.AsyncClient, messages: list[dict], **extra +) -> tuple[list[dict], list[int]]: + """Stream `/v1/chat/completions` and return the choices that carry + `token_ids` along with the prompt token IDs. + + The leading role only chunk has no `token_ids` and no counterpart on + `/inference/v1/generate`, so it is dropped. + """ + choices: list[dict] = [] + prompt_token_ids: list[int] | None = None + async with client.stream( + "POST", + "/v1/chat/completions", + json={ + "model": MODEL, + "messages": messages, + "temperature": 0, + "max_tokens": 128, + "return_token_ids": True, + "stream": True, + **extra, + }, + ) as resp: + assert resp.status_code == 200, await resp.aread() + async for line in resp.aiter_lines(): + if not line.startswith("data: "): + continue + data = line[len("data: ") :] + if data == "[DONE]": + break + chunk = json.loads(data) + assert "error" not in chunk, chunk + if chunk.get("prompt_token_ids") is not None: + prompt_token_ids = chunk["prompt_token_ids"] + choices.extend( + ch for ch in chunk["choices"] if ch.get("token_ids") is not None + ) + assert prompt_token_ids is not None + return choices, prompt_token_ids + + +def _tool_call_id_shape(tool_call_id: str | None) -> str | None: + """Drop the random suffix so IDs compare by format, e.g. `chatcmpl-tool-`. + Deterministic IDs such as kimi_k2's `functions.:` are kept + whole.""" + if tool_call_id is None: + return None + return re.sub(r"[0-9a-f]{16}$", "", tool_call_id) + + +async def _run_stream_parity_case( + client: httpx.AsyncClient, + messages: list[dict], + finish_only_tail: bool = False, + **extra, +) -> tuple[dict, list[dict]]: + """Stream the coupled request, replay its chunks through the streaming + derender endpoint and assert the assembled messages agree. Returns the + assembled coupled message and the coupled choices. + + Derender gets the engine's finish reason, since `"tool_calls"` is a chat + level rewrite of `"stop"` that derender has to reapply itself. With + `finish_only_tail` it goes on a trailing chunk with no tokens. + """ + coupled_choices, prompt_token_ids = await _coupled_stream(client, messages, **extra) + finish_reason = coupled_choices[-1]["finish_reason"] + disagg = await stream_chat_derender( + client, + [tid for ch in coupled_choices for tid in ch["token_ids"]], + [len(ch["token_ids"]) for ch in coupled_choices] + + ([0] if finish_only_tail else []), + {"model": MODEL, "messages": messages, **extra}, + len(prompt_token_ids), + prompt_token_ids, + finish_reason="stop" if finish_reason == "tool_calls" else finish_reason, + ) + coupled = assemble_stream(coupled_choices) + for message in (coupled, disagg): + for tc in message["tool_calls"]: + tc["id"] = _tool_call_id_shape(tc["id"]) + assert disagg == coupled + return coupled, coupled_choices + + +# --------------------------------------------------------------------------- +# Streaming parity cases +# --------------------------------------------------------------------------- + + +@pytest.mark.asyncio +async def test_stream_parity_plain(client): + """Plain streamed detokenization parity.""" + messages = [ + {"role": "user", "content": "What is 2+2? Answer in one short sentence."} + ] + await _run_stream_parity_case(client, messages) + + +@pytest.mark.asyncio +async def test_stream_parity_reasoning(client): + """Streamed reasoning/content split parity.""" + messages = [{"role": "user", "content": "What is 17 times 23? Think it through."}] + coupled, _ = await _run_stream_parity_case( + client, messages, include_reasoning=True, max_tokens=256 + ) + + if not coupled["reasoning"]: + pytest.skip("Model did not emit a block") + + +@pytest.mark.asyncio +async def test_stream_parity_tool_call(client): + """Streamed tool call parity under `tool_choice="auto"`, where a tool + call also rewrites `finish_reason` to `"tool_calls"`.""" + messages = [{"role": "user", "content": "What's the weather in Paris?"}] + coupled, _ = await _run_stream_parity_case( + client, messages, tools=TOOLS, tool_choice="auto", max_tokens=512 + ) + + if not coupled["tool_calls"]: + pytest.skip("Model did not emit a tool call") + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "chunking", ["engine_steps", "multi_token", "finish_only_tail"] +) +async def test_stream_parity_reasoning_and_tool_call(client, chunking): + """Streamed reasoning + forced tool call parity across chunk shapes. + + `multi_token` batches tokens per chunk with `stream_interval` the way + speculative decoding does, so replay has to reuse the original chunk + boundaries. `finish_only_tail` sends the finish reason on a chunk with no + tokens, which must still flush buffered tool call arguments. + """ + messages = [{"role": "user", "content": "What's the weather in Paris?"}] + extra: dict = { + "tools": TOOLS, + "tool_choice": FORCE_WEATHER_TOOL, + "include_reasoning": True, + "max_tokens": 512, + } + if chunking == "multi_token": + extra["stream_interval"] = 4 + coupled, coupled_choices = await _run_stream_parity_case( + client, + messages, + finish_only_tail=chunking == "finish_only_tail", + **extra, + ) + + if chunking == "multi_token": + assert any(len(ch["token_ids"]) > 1 for ch in coupled_choices) + if not (coupled["reasoning"] and coupled["tool_calls"]): + pytest.skip("Model did not emit both a block and a tool call") diff --git a/tests/entrypoints/scale_out/derender/test_derender_stream.py b/tests/entrypoints/scale_out/derender/test_derender_stream.py index e7a2e3b856..4e325db32c 100644 --- a/tests/entrypoints/scale_out/derender/test_derender_stream.py +++ b/tests/entrypoints/scale_out/derender/test_derender_stream.py @@ -5,24 +5,24 @@ Tests are split into two layers: -1. Unit tests (no server): covers ``_detokenize_delta`` correctness - (chunked == one-shot) and ``derender_completion_stream`` / - ``derender_chat_stream`` logic via a real tokenizer on a tiny model. +1. Unit tests (no server): covers `_detokenize_delta` correctness + (chunked == one-shot) and `derender_completion_stream` / + `derender_chat_stream` logic via a real tokenizer on a tiny model. The parser path is covered both with a deterministic stub parser and - with the real ``HarmonyParser`` (skipped without ``openai_harmony``). + with the real `HarmonyParser` (skipped without `openai_harmony`). 2. Integration tests (require a running render server): covers the full HTTP round-trip through the streaming endpoint. Marked with - ``@pytest.mark.asyncio`` and gated by the ``server`` / ``client`` - fixtures from the sibling ``test_derender.py``. + `@pytest.mark.asyncio` and gated by the `server` / `client` + fixtures from the sibling `test_derender.py`. """ import json -from collections.abc import Callable import pytest import pytest_asyncio +from tests.entrypoints.scale_out.derender.utils import stream_chat_derender from vllm.entrypoints.generate.base.protocol import ( DeltaFunctionCall, DeltaMessage, @@ -356,7 +356,7 @@ class TestDetokenizeDelta: """prev_tokens must not grow with the number of chunks (bounded transport). Guards that the carried decode window is a small constant - tail, so cumulative ``stream_state`` transport is O(n) and not O(n^2). + tail, so cumulative `stream_state` transport is O(n) and not O(n^2). """ token_ids = tokenizer.encode( "a reasonably long ascii stream of tokens used to exercise the " @@ -1616,86 +1616,6 @@ async def _render_parser_chat(client, messages: list[dict]) -> dict: return resp.json() -async def _stream_chat_derender( - client, - output_ids: list[int], - chunk_sizes: list[int], - chat_request: dict, - prompt_tokens: int, - prompt_token_ids: list[int], - on_chunk: Callable[[list[dict]], None] | None = None, -) -> dict: - """Feed `output_ids` through the streaming chat derender endpoint in - the given `chunk_sizes`, threading `stream_state` across calls and - return the assembled message. - - If `on_chunk` is given, it is called after every chunk with a - snapshot (deep copy) of the `tool_calls` accumulator so far, letting - callers assert properties of the intermediate deltas (e.g. monotonic - argument growth) rather than only the final assembled result. - """ - state = None - content = "" - reasoning = "" - tool_calls: list[dict] = [] - finish_reason = None - - pos = 0 - for i, size in enumerate(chunk_sizes): - tids = output_ids[pos : pos + size] - pos += size - is_last = i == len(chunk_sizes) - 1 - resp = await client.post( - "/v1/chat/completions/derender", - json={ - "stream": True, - "model": PARSER_MODEL, - "generate_chunk": { - "request_id": "stream-test", - "choices": [ - { - "index": 0, - "token_ids": tids, - "finish_reason": "stop" if is_last else None, - } - ], - }, - "stream_state": state, - "prompt_tokens": prompt_tokens, - "prompt_token_ids": prompt_token_ids, - "chat_request": chat_request, - }, - ) - assert resp.status_code == 200, resp.text - data = resp.json() - state = data["stream_state"] - delta = data["chunk"]["choices"][0]["delta"] - content += delta.get("content") or "" - reasoning += delta.get("reasoning") or "" - for tc in delta.get("tool_calls") or []: - idx = tc["index"] - while len(tool_calls) <= idx: - tool_calls.append({"id": None, "name": None, "arguments": ""}) - if tc.get("id"): - tool_calls[idx]["id"] = tc["id"] - fn = tc.get("function") or {} - if fn.get("name"): - tool_calls[idx]["name"] = fn["name"] - if fn.get("arguments"): - tool_calls[idx]["arguments"] += fn["arguments"] - if is_last: - finish_reason = data["chunk"]["choices"][0]["finish_reason"] - if on_chunk is not None: - on_chunk([dict(tc) for tc in tool_calls]) - - return { - "content": content or None, - "reasoning": reasoning or None, - "tool_calls": tool_calls, - "finish_reason": finish_reason, - } - - @pytest.mark.asyncio @pytest.mark.parametrize( "chunking", @@ -1755,7 +1675,7 @@ async def test_stream_parsed_matches_batch_reasoning( "triples": [3] * (n // 3) + [n % 3], }[chunking] chunk_sizes = [c for c in chunk_sizes if c > 0] - streamed = await _stream_chat_derender( + streamed = await stream_chat_derender( parser_client, output_ids, chunk_sizes, @@ -1823,7 +1743,7 @@ async def test_stream_parsed_matches_batch_tool_call(parser_client, parser_token if tool_calls and tool_calls[0]["arguments"]: _snapshots.append(tool_calls[0]["arguments"]) - streamed = await _stream_chat_derender( + streamed = await stream_chat_derender( parser_client, output_ids, chunk_sizes, @@ -1868,7 +1788,7 @@ async def test_stream_parsed_cjk_across_chunk_boundaries( "messages": messages, "include_reasoning": True, } - streamed = await _stream_chat_derender( + streamed = await stream_chat_derender( parser_client, output_ids, [1] * len(output_ids), @@ -1896,7 +1816,7 @@ async def test_stream_parsed_include_reasoning_false(parser_client, parser_token "messages": messages, "include_reasoning": False, } - streamed = await _stream_chat_derender( + streamed = await stream_chat_derender( parser_client, output_ids, [1] * len(output_ids), diff --git a/tests/entrypoints/scale_out/derender/utils.py b/tests/entrypoints/scale_out/derender/utils.py new file mode 100644 index 0000000000..d8128353bf --- /dev/null +++ b/tests/entrypoints/scale_out/derender/utils.py @@ -0,0 +1,103 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project +"""Helpers shared by the streaming chat derender tests.""" + +from collections.abc import Callable + +import httpx + + +def assemble_stream(choices: list[dict]) -> dict: + """Fold streamed chat choices into one message. + + A tool call index must keep the first ID it was given. On the derender + side that is what pinning the IDs in `stream_state` guarantees. + """ + content = "" + reasoning = "" + tool_calls: list[dict] = [] + finish_reason = None + for choice in choices: + delta = choice["delta"] + content += delta.get("content") or "" + reasoning += delta.get("reasoning") or "" + for tc in delta.get("tool_calls") or []: + idx = tc["index"] + while len(tool_calls) <= idx: + tool_calls.append({"id": None, "name": None, "arguments": ""}) + if tc.get("id"): + assert tool_calls[idx]["id"] in (None, tc["id"]), ( + f"tool call {idx} changed ID mid stream" + ) + tool_calls[idx]["id"] = tc["id"] + fn = tc.get("function") or {} + if fn.get("name"): + tool_calls[idx]["name"] = fn["name"] + tool_calls[idx]["arguments"] += fn.get("arguments") or "" + finish_reason = choice.get("finish_reason") or finish_reason + return { + "content": content or None, + "reasoning": reasoning or None, + "tool_calls": tool_calls, + "finish_reason": finish_reason, + } + + +async def stream_chat_derender( + client: httpx.AsyncClient, + output_ids: list[int], + chunk_sizes: list[int], + chat_request: dict, + prompt_tokens: int, + prompt_token_ids: list[int], + finish_reason: str = "stop", + on_chunk: Callable[[list[dict]], None] | None = None, +) -> dict: + """Feed `output_ids` through the streaming chat derender endpoint in + the given `chunk_sizes`, threading `stream_state` across calls and + return the assembled message. + + `finish_reason` goes on the last chunk, so a trailing size of 0 sends it + on a chunk with no tokens. + + If `on_chunk` is given, it is called after every chunk with the + `tool_calls` assembled so far, letting callers assert properties of the + intermediate deltas (e.g. monotonic argument growth) rather than only + the final assembled result. + """ + state = None + choices: list[dict] = [] + pos = 0 + for i, size in enumerate(chunk_sizes): + tids = output_ids[pos : pos + size] + pos += size + is_last = i == len(chunk_sizes) - 1 + resp = await client.post( + "/v1/chat/completions/derender", + json={ + "stream": True, + "model": chat_request["model"], + "generate_chunk": { + "request_id": "stream-test", + "choices": [ + { + "index": 0, + "token_ids": tids, + "finish_reason": finish_reason if is_last else None, + } + ], + }, + "stream_state": state, + "prompt_tokens": prompt_tokens, + "prompt_token_ids": prompt_token_ids, + "chat_request": chat_request, + }, + ) + assert resp.status_code == 200, resp.text + data = resp.json() + state = data["stream_state"] + choices.extend(data["chunk"]["choices"]) + if on_chunk is not None: + on_chunk(assemble_stream(choices)["tool_calls"]) + + return assemble_stream(choices) diff --git a/vllm/entrypoints/scale_out/token_in_token_out/protocol.py b/vllm/entrypoints/scale_out/token_in_token_out/protocol.py index 088ec4d293..fb2b938723 100644 --- a/vllm/entrypoints/scale_out/token_in_token_out/protocol.py +++ b/vllm/entrypoints/scale_out/token_in_token_out/protocol.py @@ -602,6 +602,7 @@ class DerenderChatStreamRequest(BaseModel): the client carried ``stream_state``. """ + # --8<-- [start:derender-chat-stream-request] stream: Literal[True] model: str | None = None @@ -629,6 +630,7 @@ class DerenderChatStreamRequest(BaseModel): chat_request: ChatCompletionRequest | None = None """The original (post adjust_request) ChatCompletionRequest from /render.""" + # --8<-- [end:derender-chat-stream-request] class DerenderCompletionStreamRequest(BaseModel): @@ -639,6 +641,7 @@ class DerenderCompletionStreamRequest(BaseModel): returns the derendered chunk plus updated state. """ + # --8<-- [start:derender-completion-stream-request] stream: Literal[True] model: str | None = None @@ -653,6 +656,7 @@ class DerenderCompletionStreamRequest(BaseModel): completion_request: CompletionRequest | None = None """The original (post adjust_request) CompletionRequest from /render.""" + # --8<-- [end:derender-completion-stream-request] class DerenderChatStreamResponse(BaseModel):