Skip to content

build: upgrade mamba-ssm to 2.3.2.post1, causal-conv1d to 1.6.2.post1 - #3764

Merged
ko3n1g merged 1 commit into
mainfrom
ko3n1g/build/upgrade-mamba-ssm-causal-conv1d
May 9, 2026
Merged

build: upgrade mamba-ssm to 2.3.2.post1, causal-conv1d to 1.6.2.post1#3764
ko3n1g merged 1 commit into
mainfrom
ko3n1g/build/upgrade-mamba-ssm-causal-conv1d

Conversation

@ko3n1g

@ko3n1g ko3n1g commented May 9, 2026

Copy link
Copy Markdown
Contributor
Claude summary

What

Refresh uv.lock to pick up the latest patch releases of mamba-ssm and causal-conv1d (constrained transitively to ~=2.2 / ~=1.5 via the 3rdparty/Megatron-LM submodule).

Package Before After
mamba-ssm 2.3.1 2.3.2.post1
causal-conv1d 1.6.1 1.6.2.post1

Why

Stay current on upstream patch releases of the SSM toolchain. pyproject.toml is unchanged.

How it was generated

uv lock --upgrade-package mamba-ssm --upgrade-package causal-conv1d

Test plan

  • CI green

Signed-off-by: oliver könig <okoenig@nvidia.com>
@claude

claude Bot commented May 9, 2026

Copy link
Copy Markdown
Contributor

Light Review: uv.lock refresh for mamba-ssm / causal-conv1d

Observations

  1. New transitive dependencies - mamba-ssm 2.3.2.post1 pulls in three new packages not in the prior lockfile: quack-kernels (0.4.1), tilelang (0.1.8, ~43 MB wheel), and torch-c-dlpack-ext (0.1.5). tilelang also pulls in z3-solver (~29 MB). These inflate the [ssm] optional extra significantly. Worth confirming this is expected upstream and that the image build time/size impact is acceptable.

  2. apache-tvm-ffi downgrade 0.1.11 to 0.1.9 - the resolver chose an older version to satisfy the combined constraints of tilelang, quack-kernels, and flashinfer-python. If flashinfer-python is compatible with 0.1.9 (it does not pin a lower bound in the lockfile), this should be fine, but the silent downgrade is worth double-checking against flashinfer CI.

  3. No pyproject.toml changes - confirmed. The [ssm] extra still uses unpinned mamba-ssm and causal-conv1d, so the lock refresh is the correct mechanism.

  4. No no-build-isolation-package additions needed - the new transitive deps (quack-kernels, tilelang, torch-c-dlpack-ext) all ship pre-built wheels, so no build-isolation entry is required.

No bugs, typos, or missing tests found. Straightforward lockfile refresh.

Suggested test cases

No perf tests impacted.

Since these are optional [ssm] extras, the relevant CI validation is:

  • uv sync --locked --extra ssm succeeds in the CI container
  • Any existing SSM/Mamba/NemotronH unit or functional tests pass (these exercise the mamba-ssm import path)

@ko3n1g
ko3n1g merged commit 7b7a5b8 into main May 9, 2026
23 checks passed
@ko3n1g
ko3n1g deleted the ko3n1g/build/upgrade-mamba-ssm-causal-conv1d branch May 9, 2026 13:15
gautham-kollu pushed a commit that referenced this pull request May 12, 2026
vasunvidia pushed a commit to vasunvidia/Megatron-Bridge that referenced this pull request Jun 10, 2026
…NVIDIA-NeMo#3764)

Signed-off-by: oliver könig <okoenig@nvidia.com>
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.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