name = vLLM
pull_request = https://github.com/vllm-project/vllm/pull/48792
commit = ccd8197ae9818607cb7e904ffbb83a7ce0fe31fa (refs/pull/48792/head)
upstream = https://github.com/vllm-project/vllm
license = Apache-2.0
license_file = LICENSE
copyright = Copyright contributors to the vLLM project

description =
    The Gated DeltaNet ReplaySSM standard-decode Triton kernel from vLLM PR
    #48792 ("ReplaySSM: cache SSM inputs for faster Gated DeltaNet standard
    decode"), which closed unmerged when its fork was deleted. Released vLLM ships
    ReplaySSM for Mamba2 only; unsloth_zoo.vllm_gdn_replayssm backports the GDN
    path onto released vLLM at runtime and launches this kernel.

exported_api =
    gdn_replayssm_decode.fused_recurrent_gated_delta_rule_replayssm

modifications =
    - gdn_replayssm_decode.py is vllm/third_party/flash_linear_attention/ops/
      fused_recurrent_replayssm.py from the commit above with one change, marked
      "Modified by Unsloth": the launch-config lookup imported from
      vllm.model_executor.layers.mamba.ops.replayssm_config is inlined (released
      vLLM has no "gdn_decode" entry), with the same values as the PR, and the
      kernel imports triton directly instead of vllm.triton_utils.
    - The runtime integration (state page, metadata builder, layer dispatch) is
      Unsloth's own code in unsloth_zoo/vllm_gdn_replayssm.py, ported from the
      same PR.

file_count = 1 python module + __init__.py + this MANIFEST + LICENSE

vendored_by = Unsloth (unsloth_zoo/_vendored/vllm)
