MInf: hand compact prompt ids and media tensors to the payload stager by lauradang · Pull Request #20 · tdene/Megatron-LM

lauradang · 2026-09-18T03:29:40Z

What does this PR do?

Stacked on NVIDIA#7015 (tde/ledger_capture), which is this PR's base branch.

Token capture of multimodal requests needs two things the offloaded payload did not carry:

compact_prompt_token_ids: the pre-expansion prompt the chat endpoint tokenized. A RequestPromptPreparer runs before add_request, so it can only splice a previous turn in compact space; splicing the expanded previous turn makes _build_vlm_request expand it a second time and reject the request on the placeholder count. The preparer docstring now states the compact-space contract, and a completions test pins that failure.
media_tensors: host copies of the vision encoder's inputs (imgs as packed patches, imgs_sizes, num_frames / num_tiles), so the consumer can train on exactly the pixels the policy generated against. They live on DynamicInferenceRequest.media_tensors, survive checkpoint() and merge(), and are dropped from the wire.

Both fields are None for text-only requests. Consumers: lauradang/Gym and lauradang/RL PRs stacked on NVIDIA-NeMo/Gym#2823 and NVIDIA-NeMo/RL#4129.

Pre-checks

Unit tests added (test_inference_request.py, test_multimodal_entrypoints.py, request lifecycle policy table)
Typing on new fields
black and isort run on the changed files

🤖 Generated with Claude Code

Signed-off-by: Cao Yuhang <caoyuhang@fwerkor.com> Signed-off-by: Castronaut <3116935098@qq.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

…819] (NVIDIA#7024) Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>

Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>

Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com>

Signed-off-by: Tom Long <tolong@nvidia.com>

Signed-off-by: Yaniv Galron <ygalron@nvidia.com>

Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com>

…est (NVIDIA#7041) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

Signed-off-by: wdykas <wdykas@nvidia.com>

Signed-off-by: Lawrence McAfee <lmcafee@nvidia.com>

Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

Signed-off-by: Chase Block <cblock@nvidia.com> Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com> Signed-off-by: Ravi Ghadia <rghadia@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Co-authored-by: Siddhartha Raman Sundara Raman <sraman@nvidia.com> Co-authored-by: Sudhakar Singh <sudhakars@nvidia.com> Co-authored-by: Ravi Ghadia <rghadia@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: Fei Wu <33940270+YangFei1990@users.noreply.github.com> Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com>

Signed-off-by: William Dykas <wdykas@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr1992@gmail.com> Co-authored-by: shanmugamr1992 <shanmugamr1992@gmail.com>

…erminism (NVIDIA#7069) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

Signed-off-by: Ajay Balasa <abalasa@nvidia.com>

…sformerConfig (NVIDIA#5756) Signed-off-by: Rui Zhu <rui.zhu.rz399@yale.edu> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

…biencoder_args (NVIDIA#6975) Signed-off-by: Anai-Guo <antai12232931@outlook.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

Signed-off-by: Tom Long <tolong@nvidia.com> Signed-off-by: ilml <tolong@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Signed-off-by: Zhiyu Li <zhiyul@nvidia.com>

Signed-off-by: Haoran Zhang <haoranz@nvidia.com>

Signed-off-by: Jack Chang <jianbinc@nvidia.com>

Signed-off-by: jeffnvidia <jmahou@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

Signed-off-by: dimapihtar <dpykhtar@nvidia.com> Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>

Signed-off-by: Deyu Fu <deyuf@nvidia.com> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com>

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

Signed-off-by: Faradawn Yang <73060648+faradawn@users.noreply.github.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com>

…IA#6956) Signed-off-by: Prajwal Singhania <psinghania@nvidia.com>

Signed-off-by: wdykas <wdykas@nvidia.com>

…_chunked_prefill (NVIDIA#7334) Signed-off-by: Charlie Truong <chtruong@nvidia.com>

Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>

…7293) Signed-off-by: Shiqing Fan <shiqingf@nvidia.com>

Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com>

Signed-off-by: janEbert <janpabloe@nvidia.com>

Signed-off-by: tongliu <tongliu@nvidia.com>

Resolve conflicts with main: - inference_request.py: combine main's prompt-view drop (prompt, compact, remaining) with payload-offload field stripping and stage metadata. - dynamic_engine.py: apply payload staging on main's flat-request path. - test_inference_request.py: merge the parametrized return_prompt_tokens cases. - multimodal/utils.py: removed on main by NVIDIA#7326; drop the einops guard. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

The per-request dict that rides with a SUBMIT_REQUEST to the engine's payload stager and prompt preparer was called request_metadata. That name already denotes the per-request sampling tensors on DynamicInferenceContext (request_metadata, active_request_metadata, request_metadata_types), and it sat in a wire list whose other fields are request metadata as well. Rename it end to end: REST body key, InferenceClient kwargs, coordinator handler, engine, and the DynamicInferenceRequest field. The sampling-tensor names are unchanged. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

Ship the chat-template render of the conversation through its last assistant message (template_prefix_token_ids) and the EOS id in offload_params instead of a precomputed splice boundary. The engine-side RequestPromptPreparer (NeMo RL) applies its own replace_prefix_tokens, so Megatron and NeMo RL no longer run two different boundary algorithms on one request. _replace_prefix_tokens is restored to the main version and keeps serving clients that echo prior-turn token ids and install no preparer. Remove required_prefix_token_ids: the preparer resolves the exact prefix itself, so the endpoint no longer needs to accept one. Decide the prefix source once, up front: token ids on the last assistant message (prefix replaced here) or offload_params (prefix replaced in the engine). Supplying both is rejected with 400. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

Follow the engine API after the merge with main: send flat requests to the coordinator, read step_modern()['finished_requests'], accept the offload_params keyword in add_request stubs, and unpack the 5-field SUBMIT_REQUEST metadata frame. Classify the new offload_params, payload_offloaded and payload_stage_metadata request fields in the lifecycle policy table. Add class-level None defaults for payload_stager and prompt_preparer so engines built via __new__ in tests are readable. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

- Use the module logger instead of the root logger when a prompt preparer raises. - Pack the prepared prompt and params inside the try so an unserializable preparer result falls back to the original frame stamped with the preparation error, failing the request as PromptPreparationError instead of exiting MP rank 0. - Extract _apply_prompt_preparation_error and call it from both add_request branches. - Release VLM media registered by _build_vlm_request in _handle_failed_request so a request that fails before admission does not leak it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

C12: move offload_params out of the metadata frame into a fifth client frame that the coordinator forwards verbatim (never decoded), so the frame it unpacks and repacks per request stays bounded; the engine reads it from its own frame at admission and the preparer rewrites only the prompt and offload frames. C11: schedule_requests logs a warning and drops a SUBMIT_REQUEST with the wrong frame or metadata-field count instead of raising, so one version-skewed client cannot take every MP rank down; the skip is collective because all ranks see the same broadcast list. C13: _prepare_submit_request_message returns the message untouched, without decoding any frame, when the offload frame is msgpack nil, since the preparer only has work when the client sent params. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

…contract - serialize(): an offloaded reply never carries the prompt tensors (the stager holds the prompt ids); otherwise they stay opt-in via return_prompt_tokens and are dropped when sampling_params is None, as on main. - _send_requests_to_coordinator(): build FinishedRequestRecord once per completed request and share it between the ledger and _serialize_finished_request(). - OffloadedRequestPayload.from_request(): document the prompt_tokens host-copy side effect. - payload_offloaded / payload_stage_metadata: reword the field comment; they are wire-only keys kept for deserialize(). - checkpoint()/merge(): share offload_params instead of deep-copying; nothing mutates it in place. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

The finished-request rename left two references to the old merged record in the GPU-only offload test. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

… in offload_params E19: /v1/completions now collects each reply's payload_stage_metadata, rejects conflicting values across a batch, and merges it into the top-level response body with the same reserved-field check as /v1/chat/completions; the shared logic lives in endpoints/common.py (collect_stage_metadata, attach_stage_metadata). The endpoint also accepts and forwards offload_params, which it previously ignored. E20: both endpoints reject client-supplied offload_params whose top-level keys start with '_' (HTTP 400) via endpoints.common.validate_offload_params, so the engine-owned _request_prompt_preparation_error field cannot be forged from a request body. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

Signed-off-by: Laura Dang <laurad@nvidia.com>

…load stager OffloadedRequestPayload gains compact_prompt_token_ids (the pre-expansion prompt the endpoint tokenized) and media_tensors (host copies of the vision encoder's inputs: packed-patch imgs, imgs_sizes, num_frames / num_tiles). Both are None for text-only requests. The tensors live on DynamicInferenceRequest.media_tensors, survive checkpoint() and merge(), and are dropped from the wire. Without the compact form an external RequestPromptPreparer can only splice the expanded previous turn into the compact new render, which _build_vlm_request then expands a second time and rejects on the placeholder count. The preparer docstring now states the compact-space contract, and a completions test pins the admission failure an already-expanded prefix produces. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

lauradang · 2026-09-18T08:46:52Z

/claude strict-review

@mcore

- ChainPrefixCache now requires TQTokenSource.fetch_prefix_chains. The duck-typed fallback served only a test double and, if ever taken, would have silently collapsed the compact chain onto the expanded one (the double-expansion this PR fixes). Test doubles implement the method. - Extract the media column codec into media_field_dict / row_to_media_tensors, matching the module's _row_to_* decoders, so stage_media / fetch_media are one-line delegations and the codec round-trips without a TransferQueue. - Fix comments still describing the terminal-row design: media is staged as per-call deltas and concatenated along the chain. Correct the _media_mismatch docstring (the summary carries no embedding count). Reconcile the design doc's "outside the token rows" / "ride the call row" wording and name the unpinned tdene/Megatron-LM#20 and Gym staging/media.py dependencies. Sort the tq_token_sink import block. - Tests: slice_media_tensors integrity guards (no geometry, non-dividing patch count, mid-patch boundary), media_columns_missing rejection after a swallowed stage_media failure, codec round-trip (f32 / bf16 / video / tiled), Gym extras-key and adapter-attribute parity, text runs register no media columns, expanded-space mask/logprob alignment, and an @mcore pin guard asserting OffloadedRequestPayload exposes media_tensors and compact_prompt_token_ids (red until the Megatron-Bridge pointer includes tdene/Megatron-LM#20). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

lauradang · 2026-09-21T17:14:28Z

Superseded by NVIDIA#7557, which carries the same commit rebased onto NVIDIA main now that NVIDIA#7015 has merged.

@mcore

- ChainPrefixCache now requires TQTokenSource.fetch_prefix_chains. The duck-typed fallback served only a test double and, if ever taken, would have silently collapsed the compact chain onto the expanded one (the double-expansion this PR fixes). Test doubles implement the method. - Extract the media column codec into media_field_dict / row_to_media_tensors, matching the module's _row_to_* decoders, so stage_media / fetch_media are one-line delegations and the codec round-trips without a TransferQueue. - Fix comments still describing the terminal-row design: media is staged as per-call deltas and concatenated along the chain. Correct the _media_mismatch docstring (the summary carries no embedding count). Reconcile the design doc's "outside the token rows" / "ride the call row" wording and name the unpinned tdene/Megatron-LM#20 and Gym staging/media.py dependencies. Sort the tq_token_sink import block. - Tests: slice_media_tensors integrity guards (no geometry, non-dividing patch count, mid-patch boundary), media_columns_missing rejection after a swallowed stage_media failure, codec round-trip (f32 / bf16 / video / tiled), Gym extras-key and adapter-attribute parity, text runs register no media columns, expanded-space mask/logprob alignment, and an @mcore pin guard asserting OffloadedRequestPayload exposes media_tensors and compact_prompt_token_ids (red until the Megatron-Bridge pointer includes tdene/Megatron-LM#20). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

fwerkor and others added 30 commits September 3, 2026 11:49

Fix TP=1 tensor-parallel gather autograd view (NVIDIA#6657)

1e7598c

Signed-off-by: Cao Yuhang <caoyuhang@fwerkor.com> Signed-off-by: Castronaut <3116935098@qq.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

Add per-module MFSDP prefetch budgets (NVIDIA#6905)

1030379

Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

Add flag to disable KV cache scale clamp to ModelOpt export [OMNIML-5…

da533c0

…819] (NVIDIA#7024) Signed-off-by: Jennifer Chen <jennifchen@nvidia.com>

Add HybridModel architecture guidance (NVIDIA#6938)

ae4d08b

Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>

Add a payload offload mode with a staging protocol

b48ab09

Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com>

Accept required_prefix_token_ids on the endpoint

789f746

Signed-off-by: Teodor-Dumitru Ene <teodord.ene@gmail.com>

Restore the batch schema bridge for hybrid CP sub-samples (NVIDIA#6820)

a055f8f

Signed-off-by: Tom Long <tolong@nvidia.com>

Add MTP-only training mode (NVIDIA#7021)

7db926d

Signed-off-by: Yaniv Galron <ygalron@nvidia.com>

Move MTP inference code to separate mixin class (NVIDIA#7056)

4ec08fe

Signed-off-by: Keshav Santhanam <ksanthanam@nvidia.com>

Add a MFSDP v2 port of the deepseek_proxy_fsdp_ep2_fsdp2 functional t…

91149c0

…est (NVIDIA#7041) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

build: AUT-2212 move torch-memory-saver to inference extra (NVIDIA#7068)

52500b6

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

perf(refit): reduce MXFP8 conversion overhead (NVIDIA#6952)

b1233de

Signed-off-by: wdykas <wdykas@nvidia.com>

Add current async scheduling pairwise coverage and fixes (NVIDIA#7053)

8da558d

Signed-off-by: Lawrence McAfee <lmcafee@nvidia.com>

test: clarify DeepSeek proxy MFSDP v1 names (NVIDIA#7073)

13c5581

Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

test(refit): cover DSA models and enable M2N CI (NVIDIA#6998)

6572312

Signed-off-by: William Dykas <wdykas@nvidia.com> Signed-off-by: shanmugamr1992 <shanmugamr1992@gmail.com> Co-authored-by: shanmugamr1992 <shanmugamr1992@gmail.com>

Golden deepseek_proxy_mfsdp_v2_ep2 against its own output; enable det…

c171f66

…erminism (NVIDIA#7069) Signed-off-by: Jingyue Wu <jingyuew@nvidia.com>

ci: Enhance CI env to check functional tests label (NVIDIA#7072)

1bf32f7

Signed-off-by: Ajay Balasa <abalasa@nvidia.com>

chore(checkpointing): remove dead load_biencoder_checkpoint and _add_…

fce0676

…biencoder_args (NVIDIA#6975) Signed-off-by: Anai-Guo <antai12232931@outlook.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

fix(test): AUT-2220 preserve golden value precision (NVIDIA#7087)

f8d6302

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

Fix runtime CP group handling in attention for hybrid CP (NVIDIA#6821)

df4996f

Signed-off-by: Tom Long <tolong@nvidia.com> Signed-off-by: ilml <tolong@nvidia.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

fix: restore each rank's own RNG state on checkpoint load (NVIDIA#6858)

def07af

Signed-off-by: Zhiyu Li <zhiyul@nvidia.com>

Add CuTeDSL GDP chunkwise context parallelism (NVIDIA#7038)

3b5556e

Signed-off-by: Haoran Zhang <haoranz@nvidia.com>

Update copy-pr-bot.yaml [skip ci]

b1fe759

Expose MFSDP v2 custom-schedule lifecycle (NVIDIA#6484)

b3393bb

Signed-off-by: Jack Chang <jianbinc@nvidia.com>

Fix FSDP mixed-precision reset for fixed buffers (NVIDIA#6924)

c9b53d0

Signed-off-by: jeffnvidia <jmahou@nvidia.com> Signed-off-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com> Co-authored-by: svcnvidia-nemo-ci <svcnvidia-nemo-ci@nvidia.com>

add raise_on_error argument (NVIDIA#7110)

c768946

Signed-off-by: dimapihtar <dpykhtar@nvidia.com> Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>

Add native DeepSeek-V4 hybrid attention orchestration (NVIDIA#6402)

907609b

Signed-off-by: Deyu Fu <deyuf@nvidia.com> Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com> Co-authored-by: Philip Petrakian <ppetrakian@nvidia.com>

fix: AUT-2228 route partial reruns to the current attempt (NVIDIA#7142)

62fa723

Signed-off-by: svcnemo-autobot <svcnemo-autobot@nvidia.com>

faradawn and others added 20 commits September 14, 2026 17:28

Multimodal: add tokenizer path (NVIDIA#3466)

556c74d

Signed-off-by: Faradawn Yang <73060648+faradawn@users.noreply.github.com> Co-authored-by: Dmytro Pykhtar <37850217+dimapihtar@users.noreply.github.com>

[GTP] Symmetric memory registration to use custom VMM allocator (NVID…

79097b4

…IA#6956) Signed-off-by: Prajwal Singhania <psinghania@nvidia.com>

Add refit codeowners (NVIDIA#7330)

65fe89a

Signed-off-by: wdykas <wdykas@nvidia.com>

ci: Update golden value for hybrid_dynamic_inference_tp1_pp1_dp8_583m…

ad29cd0

…_chunked_prefill (NVIDIA#7334) Signed-off-by: Charlie Truong <chtruong@nvidia.com>

Add Style Guide (NVIDIA#7179)

192b18f

Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>

Fix GTP checkpoint load failing across different GTP degrees (NVIDIA#…

8349ade

…7293) Signed-off-by: Shiqing Fan <shiqingf@nvidia.com>

Add implementation for ShortcutMoE (NVIDIA#6959)

088cd1d

Signed-off-by: Jimmy Zhang <jiemingz@nvidia.com>

Do not modify MLA spec at runtime for down proj fusion (NVIDIA#7304)

42b8a69

Signed-off-by: janEbert <janpabloe@nvidia.com>

Fix MFSDP v2 parameter metadata preservation (NVIDIA#7230)

d046b0d

Signed-off-by: tongliu <tongliu@nvidia.com>

Fix stale variable name in test_payload_offload_mode

01f1247

The finished-request rename left two references to the old merged record in the GPU-only offload test. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Laura Dang <laurad@nvidia.com>

fix(inference): type prompt preparation results

1310225

Signed-off-by: Laura Dang <laurad@nvidia.com>

lauradang mentioned this pull request Sep 18, 2026

MInf: hand compact prompt ids and media tensors to the payload stager NVIDIA/Megatron-LM#7461

Closed

5 of 6 tasks

lauradang self-assigned this Sep 18, 2026

lauradang changed the base branch from tde/ledger_capture to main September 21, 2026 16:50

lauradang mentioned this pull request Sep 21, 2026

MInf: hand compact prompt ids and media tensors to the payload stager NVIDIA/Megatron-LM#7557

Merged

5 of 6 tasks

lauradang closed this Sep 21, 2026

lauradang mentioned this pull request Sep 29, 2026

feat(token-capture): carry media through Megatron capture rollouts NVIDIA-NeMo/RL#4332

Open

Navigation Menu

MInf: hand compact prompt ids and media tensors to the payload stager - #20

MInf: hand compact prompt ids and media tensors to the payload stager#20
lauradang wants to merge 119 commits into
tdene:mainfrom
lauradang:laurad/pr7015-multimodal-capture

lauradang commented Sep 18, 2026

Uh oh!

lauradang commented Sep 18, 2026

Uh oh!

lauradang commented Sep 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

20 participants

Conversation

lauradang commented Sep 18, 2026

What does this PR do?

Pre-checks

Uh oh!

lauradang commented Sep 18, 2026

Uh oh!

lauradang commented Sep 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

20 participants