feat: Support OAI TITO v1 #502

guapisolo · 2026-01-22T01:58:31Z

This PR support several things (sorry I found it hard to split it into seperate PR.. Will do better in later PR)

Under /chat/completions endpoint with miles router enable:

Support token consistency between train & inferenc by adding trajectory manager middleware to store history samples.
Support token in for SGLang. Miles can enable token input to SGLang by setting miles_router_enable_token_input_for_chat_completions. (text model only).
Refine token out in SGLang engine, including the prompt part token out and token id (convert_ids_to_tokens sometimes has bug when the output token text include white space).
Refine agentic tool call impl.
Refine some code logic in router to make it decouple from radix tree.
Doc for refactored multi-turn & agentic
Cross turn token consistency is not supported. which means, in each turn we retokenize the OAI messages in the trajectory.
Why? Chat template is not append-only. e.g. Different model cut think part in very different timing and scenario. It makes cross turn token inherit tightly coupled with the chat template, and very hard to find a general solution.、

TODO:

Fireworks wrapper.

gemini-code-assist · 2026-01-22T02:00:18Z

Summary of Changes

Hello @guapisolo, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly overhauls the system's generation and rollout mechanisms. It introduces a new, more modular interface for defining generation functions, moving away from a monolithic approach. This refactor facilitates the integration of advanced features like multi-turn conversations and OpenAI-style tool calling, ensuring token consistency during generation. The changes also extend to the Miles router, enabling it to manage sessions and proxy these new generation types, laying the groundwork for more complex agentic behaviors.

Highlights

Refactored Rollout Function Interface: Introduced an experimental refactor of the rollout function interface, enabling more modular and flexible generation functions, activated by the MILES_EXPERIMENTAL_ROLLOUT_REFACTOR environment variable.
New Generate Function Hub: Added a new generate_hub directory containing specialized generation functions for single_turn, multi_turn, and agentic_tool_call scenarios.
OpenAI-style Tool Calling and Session Support: Implemented support for OpenAI-style chat completions with tool calling, including session management within the Miles router to maintain token consistency during generation.
Enhanced Router Functionality: The Miles router now includes session management capabilities and can proxy OpenAI-style requests, handling tokenization and record keeping for improved trajectory tracking.
Expanded Test Infrastructure: Added new test utilities (mock_sglang_server, mock_tools, uvicorn_thread_server) and comprehensive unit/integration tests to validate the new rollout and generation functionalities.

Ignored Files

Ignored by pattern: .github/workflows/** (2)
- .github/workflows/pr-test.yml
- .github/workflows/pr-test.yml.j2

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces significant refactoring to support naive OpenAI Token-In-Token-Out (TITO) for agentic workflows, particularly for multi-turn conversations with tool calls. The changes are extensive, adding new modules for generation hubs, inference rollouts, and session management, along with numerous tests. A key addition is the MILES_EXPERIMENTAL_ROLLOUT_REFACTOR feature flag to control the new logic. While the overall direction is a great improvement in modularity and functionality, I've identified a critical bug in the multi-turn logic and a few areas for improvement in the new shell script and utility functions.

gemini-code-assist · 2026-01-22T02:09:13Z

miles/rollout/generate_hub/multi_turn.py

+        if args.generate_multi_samples:
+            sample = deepcopy(input.sample)
+
+        output = await post(url, payload)
+        await update_sample_from_response(args, sample, payload=payload, output=output, update_loss_mask=True)
+
+        if args.generate_multi_samples:
+            multi_samples.append(deepcopy(sample))


There's a critical bug in the logic for generate_multi_samples. When args.generate_multi_samples is true, the sample variable is rebound to a fresh copy of input.sample on line 57. This causes the conversation history, which is maintained in the sample object, to be lost. As a result, subsequent turns in the loop will not have the correct context, as update_sample_with_tool_responses on line 75 will be updating a sample that doesn't contain the full conversation history.

To fix this, you should use a separate variable for the turn-specific sample that gets appended to multi_samples, while the main sample variable continues to hold the running state of the entire conversation.

Suggested change

if args.generate_multi_samples:

sample = deepcopy(input.sample)

output = await post(url, payload)

await update_sample_from_response(args, sample, payload=payload, output=output, update_loss_mask=True)

if args.generate_multi_samples:

multi_samples.append(deepcopy(sample))

output = await post(url, payload)

if args.generate_multi_samples:

turn_sample = deepcopy(input.sample)

await update_sample_from_response(args, turn_sample, payload=payload, output=output, update_loss_mask=True)

multi_samples.append(turn_sample)

await update_sample_from_response(args, sample, payload=payload, output=output, update_loss_mask=True)

gemini-code-assist · 2026-01-22T02:09:14Z

examples/openai_format/0.6B.sh

+sleep 3
+pkill -9 ray
+pkill -9 python


The commands to kill ray and python processes on lines 10-11 are redundant, as they are already executed on lines 7-8. The ray stop --force command should handle stopping Ray processes, and the subsequent pkill commands serve as a fallback. The second set of pkill commands is unnecessary and can be removed to simplify the script.

gemini-code-assist · 2026-01-22T02:09:14Z

miles/utils/http_utils.py


-async def post(url, payload, max_retries=60):
+# TODO may generalize the name since it now contains http DELETE/GET etc (with retries and remote-execution)
+async def post(url, payload, max_retries=60, action="post"):


The function post now handles more than just POST requests (e.g., GET, DELETE). The function name should be generalized to reflect its broader capability, for instance, to request_with_retry or a similar name that better describes its function of making various types of HTTP requests with retry logic.

gemini-code-assist · 2026-01-22T02:09:14Z

miles/utils/misc.py

+
+
+# TODO may rename to `load_object` since it can be used to load things like tool_specs
 def load_function(path):


The function load_function is now used to load objects other than functions, such as tool_specs. Renaming it to load_object would more accurately reflect its usage and prevent potential confusion.

guapisolo · 2026-01-24T02:42:37Z

@yueming-yuan Do you have some time to review this PR? I think it's almost ready. Thanks!

guapisolo requested review from fzyzcjy, yueming-yuan and yushengsu-thu as code owners January 22, 2026 01:58

gemini-code-assist bot reviewed Jan 22, 2026

View reviewed changes

guapisolo force-pushed the feat/lazy_oai_tito branch from 46d7e84 to 62176cd Compare January 23, 2026 01:08

guapisolo marked this pull request as draft January 23, 2026 01:09

guapisolo force-pushed the feat/lazy_oai_tito branch 2 times, most recently from d4655d4 to 8efb357 Compare January 23, 2026 18:55

guapisolo marked this pull request as ready for review January 23, 2026 19:04

guapisolo added 12 commits January 23, 2026 19:18

move retrieve_from_text api to middleware

d63b67d

do naive tito

aaf39a3

delete useless code

dac6d7b

fix import

efb4554

test chat input id

c92fd79

refine gen

6d229ba

small upd

c190d54

fix sglang side token output for prompt part

7162d78

fix proxy path problem

b3bda26

add 4b script

538cc13

fix try catch

b6d537b

upd test

6bebb90

guapisolo force-pushed the feat/lazy_oai_tito branch from 8efb357 to 97bc803 Compare January 23, 2026 19:18

not sure whether test is good because gpu taken away

5343a51

guapisolo force-pushed the feat/lazy_oai_tito branch from 97bc803 to 5343a51 Compare January 23, 2026 20:07

guapisolo added 2 commits January 23, 2026 20:11

upd test

a85d006

submit sglang patch

29c1595

yushengsu-thu added run-ci-fsdp run-ci-megatron labels Jan 23, 2026

fix tests

8edff99

guapisolo added 2 commits January 24, 2026 01:55

fix mock sgl server output

51c1768

fix chat input ids

0e42621

upd doc

c11ecea

guapisolo changed the title ~~Support naive OAI TITO~~ feat: Support OAI TITO v1 Jan 24, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: Support OAI TITO v1 #502

feat: Support OAI TITO v1 #502

guapisolo commented Jan 22, 2026 •

edited

Loading

Uh oh!

gemini-code-assist bot commented Jan 22, 2026

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jan 22, 2026

Uh oh!

gemini-code-assist bot Jan 22, 2026

Uh oh!

gemini-code-assist bot Jan 22, 2026

Uh oh!

gemini-code-assist bot Jan 22, 2026

Uh oh!

guapisolo commented Jan 24, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants



		# TODO may rename to `load_object` since it can be used to load things like tool_specs
		def load_function(path):

feat: Support OAI TITO v1 #502

Are you sure you want to change the base?

feat: Support OAI TITO v1 #502

Conversation

guapisolo commented Jan 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gemini-code-assist bot commented Jan 22, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

guapisolo commented Jan 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

guapisolo commented Jan 22, 2026 •

edited

Loading

guapisolo commented Jan 24, 2026 •

edited

Loading