build: widen flashinfer-python pin to <0.7.0 - #4700
Merged
Merged
Conversation
The fw-base / megatron-bridge container build pulls flashinfer-python via pip (vLLM wheel), and recent vLLM versions resolve to flashinfer-python 0.6.x. The compatible-release pin '~=0.5.0' (i.e. <0.6.0) excludes those versions and creates a system / venv split that surfaces as a runtime version-match error between flashinfer and flashinfer-cubin. Widening to <0.7.0 lets uv sync pick up the 0.6.x already installed in system Python without forcing a downgrade. Signed-off-by: oliver könig <okoenig@nvidia.com>
Contributor
Author
|
/ok to test |
svcnvidia-nemo-ci
marked this pull request as draft
May 8, 2026 11:13
Contributor
|
This PR has been automatically converted to draft because all PRs must start as drafts. When you are ready for review, click Ready for Review to begin the review process. This will:
See the contribution guide for more details. |
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Resolved 338 packages with the widened flashinfer-python constraint. flashinfer-python stays at 0.5.3 (still the latest version satisfying the constraint within MCore's resolved tree). Signed-off-by: oliver könig <okoenig@nvidia.com>
chtruong814
approved these changes
May 8, 2026
ko3n1g
enabled auto-merge
May 8, 2026 12:56
Contributor
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/25556822154 |
svcnvidia-nemo-ci
added a commit
that referenced
this pull request
May 11, 2026
- isort imports in megatron/core/datasets/data_schedule.py - remove unused TEFusedDenseMLP import in gpt_layer_specs.py - widen flashinfer-python pin to <0.7.0 in pyproject.toml + uv.lock specifier (mirrors main PR #4700) — required for MBridge compat
svcnvidia-nemo-ci
added a commit
that referenced
this pull request
May 12, 2026
- isort imports in megatron/core/datasets/data_schedule.py - remove unused TEFusedDenseMLP import in gpt_layer_specs.py - widen flashinfer-python pin to <0.7.0 in pyproject.toml + uv.lock specifier (mirrors main PR #4700) — required for MBridge compat - restore moe_expert_rank_capacity_factor field to TransformerConfig (dev-only field referenced by merged code from main) - add missing MoETokenDispatcher import to tests/unit_tests/transformer/moe/test_token_dispatcher.py
svcnvidia-nemo-ci
added a commit
that referenced
this pull request
May 12, 2026
- isort imports in megatron/core/datasets/data_schedule.py - remove unused TEFusedDenseMLP import in gpt_layer_specs.py - widen flashinfer-python pin to <0.7.0 in pyproject.toml + uv.lock specifier (mirrors main PR #4700) — required for MBridge compat - restore moe_expert_rank_capacity_factor field to TransformerConfig (dev-only field referenced by merged code from main) - add missing MoETokenDispatcher import to tests/unit_tests/transformer/moe/test_token_dispatcher.py - add --offload-optimizer-states CLI arg back to arguments.py (dev-only flag referenced by merged code) - accept dense_grouped_gemm as no-op kwarg in get_mlp_module_spec_for_backend (dev API the callers still pass) - restore mlp_hyper_connection field to TransformerLayerSubmodules (dev-only field that callers still pass) - restore TEFusedDenseMLP class to transformer_engine.py (dev-only class referenced by tests/unit_tests/transformer/test_te_fused_dense_mlp_spec.py) - add missing sharded_state_dict_default import in mlp.py - add hybrid_context_parallel field to GPTDatasetConfig (main's caller passes it, dev's class was missing it)
5 tasks
jingqiny-99
pushed a commit
to jingqiny-99/Megatron-LM
that referenced
this pull request
May 13, 2026
…dev` (NVIDIA#4760) Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: Charlie Truong <chtruong@nvidia.com> Co-authored-by: oliver könig <okoenig@nvidia.com>
yhgalaxy
pushed a commit
to yhgalaxy/Megatron-LM
that referenced
this pull request
Jun 17, 2026
Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: yhgalaxy <yhgalaxy@outlook.com>
jon-barker
pushed a commit
to jon-barker/Megatron-LM
that referenced
this pull request
Jul 10, 2026
Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: Jon Barker <jbarker@aws-cmh-slurm-1-vscode-02.cm.cluster>
terminator123
pushed a commit
to 021ai/Megatron-LM
that referenced
this pull request
Aug 3, 2026
Signed-off-by: oliver könig <okoenig@nvidia.com>
svcnvidia-nemo-ci
pushed a commit
to dimapihtar/Megatron-LM
that referenced
this pull request
Aug 4, 2026
Signed-off-by: oliver könig <okoenig@nvidia.com> Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Claude summary
Problem
megatron-core[dev]andmegatron-core[lts]capflashinfer-pythonat~=0.5.0(i.e.
>=0.5.0,<0.6.0). Downstream container builds (Megatron-Bridge fw-base)install flashinfer-python via vLLM's wheel using regular
pip, which routesaround uv and resolves to flashinfer-python
0.6.x. The result is a system //opt/venvversion split that surfaces at import time as:This blocks any clean pyproject.toml-side mitigation in Megatron-Bridge —
pinning flashinfer-python to a 0.6.x version conflicts with MCore's cap.
Fix
Widen the pin to
>=0.5.0,<0.7.0in bothdevandltsextras. Permits the0.6.x range that vLLM is already pulling, without committing to a 0.7.x major
that does not yet exist.
Test plan