Skip to content

some fix#7

Open
ZhiweiYan-96 wants to merge 2 commits into
zejun/sgl_attn_backendfrom
zhiwei/radix_attn
Open

some fix#7
ZhiweiYan-96 wants to merge 2 commits into
zejun/sgl_attn_backendfrom
zhiwei/radix_attn

Conversation

@ZhiweiYan-96
Copy link
Copy Markdown
Collaborator

No description provided.

ZhiweiYan-96 pushed a commit that referenced this pull request Apr 23, 2026
* Merge 0227 (#11)

* clear schedule redundant variables

* async basic

* update

* update util

* update multi work rpc

* update

* add connector base

* add meta info

* update, finished return correct response

* request output stage1

* request output stage2 , finished

* finished ping proxy draft

* add proxy

* atom config to connector

* proxy pass

* update start loadkv

* handshake

* sess none

* fix and add sleep

* update ,transfer without crash, need to *16, need to do get_finished&scheduler

* add finished , but 3 step run will met none

* try to add decode only

* run but hang in single test(no toproxy)

* fix hang

* sperate process kv output

* pass test , need to do decode onlyt

* update, meet prepare input bu

* run success but accuracy false

* update , seems 1 seq random correct?

* update

* disable deffered out

* run gsmk8 con=64 without crash(ds5layer)

* update 0116 status to feiyue

* update

* fix proxy

* update

* update, prefill instance can do full task

* fix high con random acc issue

* remove debug

* update

* fix bench perf

* update

* update

* gsm 0.92

* update

* fix hang

* refactor: remove redundant comments and re-adjust code organization

* fix refactor

* use transfer id (#6)

* update test

* use transfer_id

* update

* use transfer_id to fix prefill decode mismatch

* Use transfer (#7)

* update test

* use transfer_id

* update

* use transfer_id to fix prefill decode mismatch

* update assert

* fix merge ,runwithout crash

* run without crash ,but meet acc issue

* update

* fix

* enable aiter log

---------

Co-authored-by: root <root@useocpm2m-097-088.amd.com>
Co-authored-by: root <root@useocpm2m-097-083.amd.com>
Co-authored-by: knitcapcat <zejwang@amd.com>

* refine code

* fix refine

* update comment

* update deferred out config

* update proxy

* delte scripst

* clean up

* rename

* fix merge

* fix non-pd dp path

* pd+dp gsm8k pass

* update readme

* move the mesh floder

* refactor

* fix refactor

* update ut

* update ut

* fix ci

* update

* format

* format

* style: apply black formatting to pass CI checks

* fix dsv3 mtp&qwen thinking ci

* fix mtp

* Fix test_scheduler 3-tuple and revert kv_transfer_params on chat path

- tests/test_scheduler.py: update TestGetNextBatchInfo to unpack the
  (is_prefill, total_tokens, num_reqs) 3-tuple returned by
  Scheduler.get_next_batch_info. This was preexisting on main but would
  also trip atom_pd CI after the merge.
- Revert kv_transfer_params plumbing on the chat-completions path to
  match the original atom_pd scope (P/D metadata only exposed on
  /v1/completions): drop the field from ChatCompletionRequest, stop
  forwarding it in chat_completions, and remove the pass-through in
  stream_chat_response / build_chat_response / create_chat_chunk.

* style: black format kv_transfer imports after rename

The mesh -> kv_transfer rename lengthened the module path and pushed
`from atom.kv_transfer.disaggregation.base import KVConnectorBase,
KVConnectorSchedulerBase` past black's 88-char line limit. Wrap the
three affected import statements so pre-checks Black job passes.

* support AINIC in mori pd

* Apply suggestion from @Copilot

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update model_runner.py

* reformat

* ruff check

* fix some scheduler

* revsert port change

* revert mla without pd

* revert mla

* fix glm5 core dump

---------

Co-authored-by: root <root@useocpm2m-097-088.amd.com>
Co-authored-by: root <root@useocpm2m-097-083.amd.com>
Co-authored-by: knitcapcat <zejwang@amd.com>
Co-authored-by: ZhangLirong-amd <Lirong.Zhang@amd.com>
Co-authored-by: ZhangLirong <lirzhang@amd.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant