refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep by akoumpa · Pull Request #2630 · NVIDIA-NeMo/Automodel

akoumpa · 2026-06-18T06:52:21Z

What does this PR do ?

Removes the deprecated BackendConfig.enable_deepep flag and switches the MoE recipes failing the 26.06 release-testing DeepEP internode-dispatch faults to the HybridEP dispatcher.

enable_deepep silently rewrote experts/dispatcher in __post_init__, and since that ran last it would override an explicitly-set dispatcher (a stale enable_deepep: true next to dispatcher: hybridep was forced back to deepep). It is now ignored with a logged warning if set — the field is retained only so an old config does not crash the kw_only dataclass — and all in-repo configs/tests are migrated to explicit dispatcher/experts.

Changelog

BackendConfig: stop honoring enable_deepep; log a warning (rank 0) if set, and no longer alter dispatcher/experts.
Migrate 25 configs off enable_deepep, preserving prior runtime behavior: true → experts: gmm + dispatcher: deepep, false → dispatcher: torch.
- 5 configs wrote experts: te alongside enable_deepep: true; the flag was silently forcing experts=gmm at runtime, so they migrate to experts: gmm (flag if te was actually intended).
Switch the failing ep_size>8 recipes to dispatcher: hybridep (Linear AM-450/473/488/490): minimax_m2.{1,5,7}_hellaswag_pp, qwen3_next_te_deepep (benchmark+finetune), glm_4.7_te_deepep (benchmark+finetune), qwen3_vl_moe_235b, step_3.5_flash_hellaswag_pp. Mirrors fix(docker): build DeepEP against the NVSHMEM wheel matching the apt runtime #2614.
Update the nemotron-omni docs example and unit tests; rewrite the enable_deepep test cases to assert the ignore+warn behavior; drop one pre-existing unused-variable (F841) in a touched test.

Before your PR is "Ready for review"

Pre checks:

Make sure you read and followed Contributor guidelines
Did you write any new necessary tests? — rewrote tests/unit_tests/moe/test_backend_config.py enable_deepep cases (49 passed locally)
Did you add or update any necessary documentation? — docs/guides/vlm/nemotron-omni

Additional Information

Mirrors fix(docker): build DeepEP against the NVSHMEM wheel matching the apt runtime #2614 (glm_4.5_air HybridEP fix).
Addresses Linear AM-450, AM-473, AM-488, AM-490 (26.06 release testing, ep_size>8 DeepEP internode dispatch faults).

…ridep enable_deepep was a deprecated BackendConfig flag whose __post_init__ conversion silently rewrote experts/dispatcher -- and because it ran last, it overrode any explicitly-set dispatcher (so `dispatcher: hybridep` next to a stale `enable_deepep: true` was forced back to deepep). Remove the conversion. enable_deepep is now ignored with a logged warning if a config still sets it; the field is retained only so an old config does not crash the kw_only dataclass. All in-repo configs/tests are migrated to explicit fields, preserving prior runtime behavior: - enable_deepep: true -> experts: gmm, dispatcher: deepep - enable_deepep: false -> dispatcher: torch - 5 configs wrote `experts: te` alongside `enable_deepep: true`; the flag was silently forcing experts=gmm at runtime, so they migrate to `experts: gmm`. Also switch the recipes hitting the 26.06 release-testing DeepEP internode dispatch faults (recv counters stuck at -1 / CPU-dispatch timeout; AM-450, AM-473, AM-488, AM-490, all ep_size>8) to the HybridEP dispatcher, mirroring #2614 (glm_4.5_air) and the gpt_oss_120b / ling_1t_sft / *_gb200 recipes already on it: minimax_m2.{1,5,7}_hellaswag_pp, qwen3_next_te_deepep (benchmark+finetune), glm_4.7_te_deepep (benchmark+finetune), qwen3_vl_moe_235b, step_3.5_flash_hellaswag_pp. Tests: rewrite the BackendConfig enable_deepep cases to assert ignore+warn, and migrate model-test BackendConfig kwargs from enable_deepep=False to dispatcher="torch". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

copy-pr-bot · 2026-06-18T06:52:24Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

github-actions · 2026-06-18T06:54:03Z

🌿 Preview your docs: https://nvidia-preview-akoumpa-refactor-remove-enable-deepep.docs.buildwithfern.com/nemo/automodel

HuiyingLi · 2026-06-18T07:08:12Z

    rms_norm: torch_fp32
    rope_fusion: false
-    enable_deepep: false
+    dispatcher: torch


Good catch. enable_deepep: false had migrated 1:1 to dispatcher: torch, but this example is ep_size: 8 (intra-node), so deepep is appropriate. Changed to experts: gmm + dispatcher: deepep in 8c83a5b.

Rule applied across these: ep_size > 8 (internode) → hybridep, ep_size <= 8 → deepep, paired with experts: gmm to match the real examples/vlm_finetune/nemotron_omni/* configs.

HuiyingLi · 2026-06-18T07:08:45Z

    rms_norm: torch_fp32
    rope_fusion: false
-    enable_deepep: false
+    dispatcher: torch


if the original one is torch can we change it to deepep/hybridep

Yes — done in 8c83a5b. This example is ep_size: 8 (intra-node), so it now uses experts: gmm + dispatcher: deepep. The rule: ep_size <= 8 → deepep, ep_size > 8 (internode) → hybridep.

HuiyingLi · 2026-06-18T07:09:22Z

  backend:
    _target_: nemo_automodel.components.models.common.utils.BackendConfig
-    enable_deepep: false
+    dispatcher: torch


Done (8c83a5b): switched to experts: gmm + dispatcher: deepep — this config is ep_size: 8 (intra-node). Applied the same to the other customizer_gpt_oss_*_full_sft* and customizer_nemotron_nano_* configs that have EP.

The single-GPU PEFT variants (customizer_gpt_oss_peft{,_packing}, ep_size: null, nproc_per_node: 1) stay on torch — no expert parallelism, so deepep/hybridep (multi-rank all-to-all) would not apply.

The enable_deepep: false -> dispatcher: torch migrations left configs that have real expert parallelism on the slow torch reference path. Switch them to a performant dispatcher per ep_size: ep_size > 8 (internode) -> hybridep, ep_size <= 8 (intra-node) -> deepep, each paired with experts: gmm to match the sibling deepep configs for these models. Single-GPU PEFT configs (ep_size=null) and the 4-layer functional test keep dispatcher: torch (no/minimal expert parallelism). Addresses review feedback from HuiyingLi on #2630. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

Pipeline 55187333 (main-mirror, eos/H100) shows ep_size>8 deepep recipes faulting in DeepEP internode dispatch (RuntimeError: DeepEP error: timeout (dispatch CPU) at internode_dispatch). On H100 (8 GPUs/node) any ep_size>8 spans nodes, so deepep internode dispatch faults; hybridep is the validated internode path (mirrors #2614). Switch deepep -> hybridep: - glm_5_hellaswag_pp, glm_5.1_hellaswag_pp (ep64; set to deepep earlier in this PR, but ep64 > 8 so they must be hybridep) - glm_5.1_lora, ling_flash_2_0_sft, minimax_m2.7_hellaswag_lora (ep32) - deepseek_v4_flash_hellaswag{,_lora,_packed_sequence} (ep32), deepseek_v4_pro_*_ep64 (ep64), hy3_preview_deepep_lora (ep64), mimo_v2_flash_hellaswag (ep32), nemotron_ultra_v3_hellaswag_peft (ep32) nemotron_ultra_v3_squad (ep64) is left as-is: it fails an unrelated check (non_pp_size must be a multiple of ep_size) from running a 16-node recipe on 1 node, not a dispatch fault. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

akoumpa · 2026-06-18T19:32:30Z

Addressed @HuiyingLi's review + a follow-up `ep_size>8` sweep

8c83a5b — the enable_deepep: false → dispatcher: torch migrations that actually have expert parallelism now use a performant dispatcher instead of the torch reference path. Rule: ep_size > 8 (internode) → hybridep, ep_size <= 8 (intra-node) → deepep, paired with experts: gmm:

ep_size 8 → deepep: nemotron-omni docs (.md/.mdx), customizer_gpt_oss_full_sft{,_chat}, customizer_nemotron_nano_full_sft{,_chat}
ep_size 4 → deepep: customizer_nemotron_nano_peft{,_packing}
ep_size 32 → hybridep: qwen3_5_moe_medpix
left on torch: customizer_gpt_oss_peft{,_packing} (ep_size: null, single GPU — no EP to dispatch)

c29d447 — scanned nemo-ci pipeline 55187333 (main-mirror, eos/H100) for other ep_size>8 recipes still on deepep. On H100 (8 GPUs/node) any ep_size>8 crosses nodes, where DeepEP internode dispatch faults (RuntimeError: DeepEP error: timeout (dispatch CPU)). Switched 13 → hybridep:

confirmed internode-dispatch fault/hang: glm_5_hellaswag_pp, glm_5.1_hellaswag_pp (both ep64 — these were set to deepep earlier in this PR), glm_5.1_lora, ling_flash_2_0_sft, minimax_m2.7_hellaswag_lora
ep>8 deepep, masked by an HF-download infra failure this run but on the same internode path: deepseek_v4_flash_hellaswag{,_lora,_packed_sequence}, deepseek_v4_pro_*_ep64 (×2), hy3_preview_deepep_lora, mimo_v2_flash_hellaswag, nemotron_ultra_v3_hellaswag_peft

Not touched: nemotron_ultra_v3_squad (ep64) fails an unrelated check (non_pp_size=8 must be a multiple of ep_size=64) — a 16-node recipe ran on 1 node; needs >=8 nodes (designed for 16), not a dispatcher change.

Without a ci section the harness defaults to 1 node (8 GPUs), which fails the "non_pp_size must be a multiple of ep_size" check (non_pp_size=8 vs ep_size=64). Add a ci section requesting 8 nodes (pp_size(1) * ep_size(64) = 64 GPUs) and assign recipe_owner adil-a, mirroring nemotron_ultra_v3_hellaswag_peft.yaml. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

Without a ci section the harness defaults to 1 node, which fails the "non_pp_size must be a multiple of ep_size" check for ep_size>8 recipes. Add ci sections (recipe_owner HuiyingLi), sizing nodes to pp_size*ep_size: - nemotron_ultra_v3_hellaswag_peft_gb200 (ep16, 4 GB200/node): 4 nodes - minimax_m3_vl_sft_ep32pp4 (ep32, pp4): 16 nodes - deepseekv3_pretrain (ep16, pp8): 16 nodes Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

step3p7_medpix_200b_ep32pp4 (ep32; surfaced failing in pipeline 55187316) and minimax_m3_vl_sft_ep32pp4 (ep32) are internode on H100 (8 GPUs/node), where DeepEP dispatch faults. hybridep is the validated internode path (mirrors #2614). Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

validate-nightly-recipes (tools/lint_example_yamls.py) requires the top-level ci section to be the last section; the ci block I added landed before the wandb section. Move it after wandb so ci is the last top-level key. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

HuiyingLi · 2026-06-19T03:58:59Z

  backend:
    _target_: nemo_automodel.components.models.common.utils.BackendConfig
-    enable_deepep: false
+    dispatcher: torch


Yes, intentional. This PEFT recipe is single-GPU (ep_size: null, dp_size: 1, nproc_per_node: 1), so there is no expert parallelism — deepep/hybridep are multi-rank all-to-all dispatchers and do not apply, so torch is correct here. (The *_full_sft* variants are ep_size: 8 and did move to deepep.)

HuiyingLi · 2026-06-19T03:59:29Z

  backend:
    _target_: nemo_automodel.components.models.common.utils.BackendConfig
-    enable_deepep: false
+    dispatcher: torch


Same — customizer_gpt_oss_peft_packing is also single-GPU (ep_size: null, nproc_per_node: 1), so torch is intentional (no EP to dispatch).

HuiyingLi · 2026-06-19T03:59:52Z

+
+ci:
+  # pp_size(1) * ep_size(16) = 16 GPUs => 4 nodes (4 GB200/node).
+  recipe_owner: HuiyingLi


Ultra is Adil?

Yes — fixed to adil-a in 70fb43e (matches nemotron_ultra_v3_squad and nemotron_ultra_v3_hellaswag_peft).

HuiyingLi · 2026-06-19T04:00:15Z

+
+ci:
+  # pp_size(8) * ep_size(16) = 128 GPUs => 16 nodes (8 H100/node).
+  recipe_owner: HuiyingLi


v3 is hemil?

Yes — fixed to hemildesai in 70fb43e (matches the deepseek_v4_* recipes).

HuiyingLi · 2026-06-19T04:00:33Z

+
+ci:
+  # pp_size(4) * ep_size(32) = 128 GPUs => 16 nodes (8 H100/node).
+  recipe_owner: HuiyingLi


m3 is Abhishree?

Yes — fixed to athitten (Abhishree) in 70fb43e.

@HuiyingLi

Per @HuiyingLi review on #2630, the ci sections should name the actual model owners rather than a single placeholder: - nemotron_ultra_v3_hellaswag_peft_gb200: -> adil-a (matches nemotron_ultra_v3_squad / _hellaswag_peft) - deepseekv3_pretrain: -> hemildesai (matches the deepseek_v4 recipes) - minimax_m3_vl_sft_ep32pp4: -> athitten Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

HuiyingLi · 2026-06-19T14:28:32Z

/ok to test 70fb43e

…o hybridep (2630)` into `r0.5.0` Cherry-pick of #2630 (50a12b7) onto r0.5.0. Release-branch resolutions: - Dropped 4 files modified by #2630 that do not exist on r0.5.0 (the models were not backported): examples/vlm_finetune/minimax_m3/minimax_m3_vl_sft_ep32pp4.yaml, tests/unit_tests/models/minimax_m3_vl/{conftest.py,test_minimax_rope_fp32.py}, tests/unit_tests/models/qwen3_5/test_qwen3_5_dense_backbone.py. - Applied #2630's enable_deepep=False -> dispatcher="torch" migration to tests/unit_tests/models/qwen3_5/test_qwen3_5_pp.py as well. #2630 did not touch this file (main migrated it earlier in #2557, not yet on r0.5.0); the stale enable_deepep=False would otherwise become a silently-ignored no-op, letting the backend default to dispatcher="deepep" on the GPU runner. The post-merge CI failures on main for 50a12b7 were pre-existing L0 unit-test suite timeouts (CPU 600s, GPU 1800s) seen on the parent commit too, not a regression introduced by this change. --- refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep (#2630) enable_deepep was a deprecated BackendConfig flag whose __post_init__ conversion silently rewrote experts/dispatcher -- and because it ran last, it overrode any explicitly-set dispatcher (so `dispatcher: hybridep` next to a stale `enable_deepep: true` was forced back to deepep). Remove the conversion. enable_deepep is now ignored with a logged warning if a config still sets it; the field is retained only so an old config does not crash the kw_only dataclass. All in-repo configs/tests are migrated to explicit fields, preserving prior runtime behavior: - enable_deepep: true -> experts: gmm, dispatcher: deepep - enable_deepep: false -> dispatcher: torch - 5 configs wrote `experts: te` alongside `enable_deepep: true`; the flag was silently forcing experts=gmm at runtime, so they migrate to `experts: gmm`. Also switch the recipes hitting the 26.06 release-testing DeepEP internode dispatch faults (recv counters stuck at -1 / CPU-dispatch timeout; AM-450, AM-473, AM-488, AM-490, all ep_size>8) to the HybridEP dispatcher, mirroring #2614 (glm_4.5_air) and the gpt_oss_120b / ling_1t_sft / *_gb200 recipes already on it. Tests: rewrite the BackendConfig enable_deepep cases to assert ignore+warn, and migrate model-test BackendConfig kwargs from enable_deepep=False to dispatcher="torch". (cherry picked from commit 50a12b7) Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…o hybridep (2630)` into `r0.5.0` Cherry-pick of #2630 (50a12b7) onto r0.5.0. Release-branch resolutions: - Dropped 4 files modified by #2630 that do not exist on r0.5.0 (the models were not backported): examples/vlm_finetune/minimax_m3/minimax_m3_vl_sft_ep32pp4.yaml, tests/unit_tests/models/minimax_m3_vl/{conftest.py,test_minimax_rope_fp32.py}, tests/unit_tests/models/qwen3_5/test_qwen3_5_dense_backbone.py. - Applied #2630's enable_deepep=False -> dispatcher="torch" migration to tests/unit_tests/models/qwen3_5/test_qwen3_5_pp.py as well. #2630 did not touch this file (main migrated it earlier in #2557, not yet on r0.5.0); the stale enable_deepep=False would otherwise become a silently-ignored no-op, letting the backend default to dispatcher="deepep" on the GPU runner. The post-merge CI failures on main for 50a12b7 were pre-existing L0 unit-test suite timeouts (CPU 600s, GPU 1800s) seen on the parent commit too, not a regression introduced by this change. --- refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep (#2630) enable_deepep was a deprecated BackendConfig flag whose __post_init__ conversion silently rewrote experts/dispatcher -- and because it ran last, it overrode any explicitly-set dispatcher (so `dispatcher: hybridep` next to a stale `enable_deepep: true` was forced back to deepep). Remove the conversion. enable_deepep is now ignored with a logged warning if a config still sets it; the field is retained only so an old config does not crash the kw_only dataclass. All in-repo configs/tests are migrated to explicit fields, preserving prior runtime behavior: - enable_deepep: true -> experts: gmm, dispatcher: deepep - enable_deepep: false -> dispatcher: torch - 5 configs wrote `experts: te` alongside `enable_deepep: true`; the flag was silently forcing experts=gmm at runtime, so they migrate to `experts: gmm`. Also switch the recipes hitting the 26.06 release-testing DeepEP internode dispatch faults (recv counters stuck at -1 / CPU-dispatch timeout; AM-450, AM-473, AM-488, AM-490, all ep_size>8) to the HybridEP dispatcher, mirroring #2614 (glm_4.5_air) and the gpt_oss_120b / ling_1t_sft / *_gb200 recipes already on it. Tests: rewrite the BackendConfig enable_deepep cases to assert ignore+warn, and migrate model-test BackendConfig kwargs from enable_deepep=False to dispatcher="torch". (cherry picked from commit 50a12b7) Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

…o hybridep (2630)` into `r0.5.0` (#2644) Cherry-pick of #2630 (50a12b7) onto r0.5.0. Release-branch resolutions: - Dropped 4 files modified by #2630 that do not exist on r0.5.0 (the models were not backported): examples/vlm_finetune/minimax_m3/minimax_m3_vl_sft_ep32pp4.yaml, tests/unit_tests/models/minimax_m3_vl/{conftest.py,test_minimax_rope_fp32.py}, tests/unit_tests/models/qwen3_5/test_qwen3_5_dense_backbone.py. - Applied #2630's enable_deepep=False -> dispatcher="torch" migration to tests/unit_tests/models/qwen3_5/test_qwen3_5_pp.py as well. #2630 did not touch this file (main migrated it earlier in #2557, not yet on r0.5.0); the stale enable_deepep=False would otherwise become a silently-ignored no-op, letting the backend default to dispatcher="deepep" on the GPU runner. The post-merge CI failures on main for 50a12b7 were pre-existing L0 unit-test suite timeouts (CPU 600s, GPU 1800s) seen on the parent commit too, not a regression introduced by this change. --- refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep (#2630) enable_deepep was a deprecated BackendConfig flag whose __post_init__ conversion silently rewrote experts/dispatcher -- and because it ran last, it overrode any explicitly-set dispatcher (so `dispatcher: hybridep` next to a stale `enable_deepep: true` was forced back to deepep). Remove the conversion. enable_deepep is now ignored with a logged warning if a config still sets it; the field is retained only so an old config does not crash the kw_only dataclass. All in-repo configs/tests are migrated to explicit fields, preserving prior runtime behavior: - enable_deepep: true -> experts: gmm, dispatcher: deepep - enable_deepep: false -> dispatcher: torch - 5 configs wrote `experts: te` alongside `enable_deepep: true`; the flag was silently forcing experts=gmm at runtime, so they migrate to `experts: gmm`. Also switch the recipes hitting the 26.06 release-testing DeepEP internode dispatch faults (recv counters stuck at -1 / CPU-dispatch timeout; AM-450, AM-473, AM-488, AM-490, all ep_size>8) to the HybridEP dispatcher, mirroring #2614 (glm_4.5_air) and the gpt_oss_120b / ling_1t_sft / *_gb200 recipes already on it. Tests: rewrite the BackendConfig enable_deepep cases to assert ignore+warn, and migrate model-test BackendConfig kwargs from enable_deepep=False to dispatcher="torch". (cherry picked from commit 50a12b7) Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>

akoumpa requested review from a team and jgerh as code owners June 18, 2026 06:52

akoumpa added the r0.5.0 Auto-cherrypick to release branch. Apply before merge; cherrypick happens after merge. label Jun 18, 2026

HuiyingLi reviewed Jun 18, 2026

View reviewed changes

akoumpa added 2 commits June 18, 2026 12:22

akoumpa added 4 commits June 18, 2026 12:36

akoumpa requested a review from HuiyingLi June 19, 2026 03:25

HuiyingLi reviewed Jun 19, 2026

View reviewed changes

HuiyingLi approved these changes Jun 19, 2026

View reviewed changes

HuiyingLi enabled auto-merge (squash) June 19, 2026 14:28

copy-pr-bot Bot temporarily deployed to nemo-ci June 19, 2026 14:28 Inactive

copy-pr-bot Bot temporarily deployed to test June 19, 2026 14:28 Inactive

copy-pr-bot Bot temporarily deployed to public June 19, 2026 14:29 Inactive

copy-pr-bot Bot temporarily deployed to public June 19, 2026 14:31 Inactive

copy-pr-bot Bot temporarily deployed to nemo-ci June 19, 2026 14:33 Inactive

HuiyingLi merged commit 50a12b7 into main Jun 19, 2026
82 checks passed

HuiyingLi deleted the akoumpa/refactor/remove-enable-deepep branch June 19, 2026 16:05

akoumpa mentioned this pull request Jun 19, 2026

cp: refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep (2630) into r0.5.0 #2644

Merged

This was referenced Jun 20, 2026

fix(model): apply dict config overrides with custom dispatch #2657

Merged

cp: feat(qwen3_5): port dense Qwen3.5 to a custom-model(2557) into r0.5.0 #2663

Merged

Uh oh!

Conversation

akoumpa commented Jun 18, 2026

What does this PR do ?

Changelog

Before your PR is "Ready for review"

Additional Information

Uh oh!

copy-pr-bot Bot commented Jun 18, 2026

Uh oh!

github-actions Bot commented Jun 18, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

akoumpa commented Jun 18, 2026

Addressed @HuiyingLi's review + a follow-up ep_size>8 sweep

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

HuiyingLi commented Jun 19, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Addressed @HuiyingLi's review + a follow-up `ep_size>8` sweep