MANIFEST.in
README.md
pyproject.toml
b12x/__init__.py
b12x.egg-info/PKG-INFO
b12x.egg-info/SOURCES.txt
b12x.egg-info/dependency_links.txt
b12x.egg-info/entry_points.txt
b12x.egg-info/requires.txt
b12x.egg-info/top_level.txt
b12x/attention/__init__.py
b12x/attention/copy_utils.py
b12x/attention/pipeline.py
b12x/attention/reference.py
b12x/attention/utils.py
b12x/attention/mla/__init__.py
b12x/attention/mla/api.py
b12x/attention/mla/kernel.py
b12x/attention/mla/reference.py
b12x/attention/mla/split.py
b12x/attention/mla/traits.py
b12x/attention/mla/workspace.py
b12x/attention/nsa_indexer/__init__.py
b12x/attention/nsa_indexer/api.py
b12x/attention/nsa_indexer/extend_kernel.py
b12x/attention/nsa_indexer/kernel.py
b12x/attention/nsa_indexer/reference.py
b12x/attention/paged/__init__.py
b12x/attention/paged/api.py
b12x/attention/paged/forward_extend_generic.py
b12x/attention/paged/forward_paged.py
b12x/attention/paged/graph_replay.py
b12x/attention/paged/merge.py
b12x/attention/paged/planner.py
b12x/attention/paged/traits.py
b12x/attention/paged/workspace.py
b12x/attention/paged/tuning/__init__.py
b12x/attention/paged/tuning/bf16.decode.bs1.py
b12x/attention/paged/tuning/bf16.decode.bs12.py
b12x/attention/paged/tuning/bf16.decode.bs128.py
b12x/attention/paged/tuning/bf16.decode.bs16.py
b12x/attention/paged/tuning/bf16.decode.bs2.py
b12x/attention/paged/tuning/bf16.decode.bs32.py
b12x/attention/paged/tuning/bf16.decode.bs4.py
b12x/attention/paged/tuning/bf16.decode.bs6.py
b12x/attention/paged/tuning/bf16.decode.bs64.py
b12x/attention/paged/tuning/bf16.decode.bs8.py
b12x/attention/paged/tuning/fp8.decode.bs1.py
b12x/attention/paged/tuning/fp8.decode.bs12.py
b12x/attention/paged/tuning/fp8.decode.bs128.py
b12x/attention/paged/tuning/fp8.decode.bs16.py
b12x/attention/paged/tuning/fp8.decode.bs2.py
b12x/attention/paged/tuning/fp8.decode.bs32.py
b12x/attention/paged/tuning/fp8.decode.bs4.py
b12x/attention/paged/tuning/fp8.decode.bs64.py
b12x/attention/paged/tuning/fp8.decode.bs8.py
b12x/attention/paged/tuning/registry.py
b12x/cute/__init__.py
b12x/cute/fp4.py
b12x/cute/runtime_patches.py
b12x/cute/utils.py
b12x/distributed/__init__.py
b12x/distributed/_cuda_ipc.py
b12x/distributed/pcie_oneshot.cu
b12x/distributed/pcie_oneshot.py
b12x/gemm/__init__.py
b12x/gemm/dense.py
b12x/integration/__init__.py
b12x/integration/attention.py
b12x/integration/mla.py
b12x/integration/nsa_indexer.py
b12x/integration/tp_moe.py
b12x/integration/triton_compact.py
b12x/integration/triton_route.py
b12x/moe/__init__.py
b12x/moe/fused/__init__.py
b12x/moe/fused/dynamic.py
b12x/moe/fused/micro.py
b12x/moe/fused/reference.py
b12x/moe/fused/relu2.py
b12x/moe/fused/silu.py
b12x/moe/fused/static.py
b12x/moe/tuning/__init__.py
b12x/moe/tuning/decode.max_active_clusters.py
b12x/moe/tuning/registry.py
b12x/quant/__init__.py
b12x/quant/expert_fp4.py
b12x/quantization/__init__.py
b12x/quantization/bf16_to_fp4_tma.py
serve/__init__.py
serve/cli.py
serve/logging.py
serve/runtime_warnings.py
serve/api/__init__.py
serve/api/server.py
serve/api/webui.py
serve/cache/__init__.py
serve/cache/kv_cache.py
serve/cache/linear_state_arena.py
serve/cache/mamba_pool.py
serve/cache/page_pool.py
serve/cache/prefix_checkpoint_cache.py
serve/cache/radix_cache.py
serve/cache/ssm_pool.py
serve/cache/tensor_arena.py
serve/engine/__init__.py
serve/engine/cuda_graph.py
serve/engine/mamba_metadata.py
serve/engine/request.py
serve/engine/runner.py
serve/engine/sampling.py
serve/engine/scheduler.py
serve/engine/serving.py
serve/engine/step_state.py
serve/kernels/__init__.py
serve/kernels/fla/__init__.py
serve/kernels/fla/chunk.py
serve/kernels/fla/chunk_delta_h.py
serve/kernels/fla/cumsum.py
serve/kernels/fla/fused_gdn_gating.py
serve/kernels/fla/fused_norm_gate.py
serve/kernels/fla/fused_recurrent.py
serve/kernels/fla/fused_sigmoid_gating_recurrent.py
serve/kernels/fla/solve_tril.py
serve/kernels/mamba/__init__.py
serve/kernels/mamba/causal_conv1d.py
serve/kernels/mamba/causal_conv1d_triton.py
serve/model/__init__.py
serve/model/attention.py
serve/model/ffn.py
serve/model/gdn.py
serve/model/layer.py
serve/model/loader.py
serve/model/ops.py
serve/model/recipe_minimax_m2.py
serve/model/recipe_qwen3_5.py
serve/model/workspaces.py
serve/tp/__init__.py
serve/tp/group.py
serve/tp/launch.py