Add Mamba-2 kernels (causal-conv1d + mamba-ssm) to slime image for Nemotron-H

[yeshsurya]

Why

The slime-pytorch-2.9-cuda12.8 curated image does not ship the Mamba CUDA kernels, so Megatron's MambaMixer aborts with ImportError: MambaSSM is not installed at model-build time. This blocks training hybrid Mamba-2 / MoE models such as Nemotron-H (model_type: nemotron_h), even though the image already contains:

• the native nemotron_h bridge in the bundled megatron.bridge (megatron/bridge/models/nemotronh/nemotron_h_bridge.py), and
• nemotron_h rollout support in sglang 0.5.11.

Kernels were the only missing piece for end-to-end Nemotron-H post-training.

What

• Dockerfile: build causal-conv1d v1.5.0.post8 and mamba-ssm v2.3.2.post1 from their pinned git tags after the apex step.
  • Built from git tags because the PyPI sdists omit csrc/*.cpp and fail the ninja build (missing and no known rule to make it).
  • mamba-ssm 2.3.2.post1 imports the modern transformers.generation.GenerateDecoderOnlyOutput symbol and guards its optional TileLang (Mamba-3 MIMO) import, so it loads cleanly on this image's transformers with no source patch.
  • --no-deps keeps the curated torch / sglang / tilelang stack pinned (the image installs a cu128-nightly tilelang for flash-linear-attention; the Mamba-2 path Nemotron-H uses does not need tilelang).
  • TORCH_CUDA_ARCH_LIST="8.0;9.0" covers A100 + H100. Kernel refs are overridable ARGs.
• smoke_test.py: assert the causal_conv1d + mamba_ssm.ops.triton kernels Megatron's MambaMixer uses import, and mamba_ssm >= 2.3.2.
• spec.yaml: advertise Mamba: "2".

Validation

The causal-conv1d 1.5.0.post8 + mamba-ssm 2.3.2.post1 pair was validated end-to-end against :7 on a cluster: both kernels import, and Megatron wires the mamba params (mamba_num_heads=64, mamba_num_groups=8, mamba_state_dim=128, spec=mamba_stack_spec). Image auto-versions via asset.yaml (version: auto).