LICENSE
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/_lib/__init__.py
b12x/_lib/compiler.py
b12x/_lib/dense_gemm.py
b12x/_lib/dense_gemm_mxfp6.py
b12x/_lib/env.py
b12x/_lib/fp6.py
b12x/_lib/gating.py
b12x/_lib/intrinsics.py
b12x/_lib/meta.py
b12x/_lib/runtime_control.py
b12x/_lib/runtime_patches.py
b12x/_lib/scratch.py
b12x/_lib/scratch_layout.py
b12x/_lib/smem.py
b12x/_lib/utils.py
b12x/_lib/quant/__init__.py
b12x/_lib/quant/mxfp8_rows.py
b12x/_lib/quant/sqg_e4m3.py
b12x/_lib/quant/sqg_fp16_d3l.py
b12x/_lib/quant/x4t_scales.py
b12x/attention/__init__.py
b12x/attention/_shared/__init__.py
b12x/attention/_shared/workspace.py
b12x/attention/_shared/contiguous/__init__.py
b12x/attention/_shared/contiguous/api.py
b12x/attention/_shared/contiguous/block_info.py
b12x/attention/_shared/contiguous/cute_dsl_utils.py
b12x/attention/_shared/contiguous/forward.py
b12x/attention/_shared/contiguous/layout_utils.py
b12x/attention/_shared/contiguous/mask.py
b12x/attention/_shared/contiguous/named_barrier.py
b12x/attention/_shared/contiguous/pack_gqa.py
b12x/attention/_shared/contiguous/seqlen_info.py
b12x/attention/_shared/contiguous/softmax.py
b12x/attention/_shared/contiguous/tile_scheduler.py
b12x/attention/_shared/cute/__init__.py
b12x/attention/_shared/cute/copy.py
b12x/attention/_shared/cute/ops.py
b12x/attention/_shared/cute/pipeline.py
b12x/attention/_shared/mla/__init__.py
b12x/attention/_shared/mla/api.py
b12x/attention/_shared/mla/compressed_api.py
b12x/attention/_shared/mla/compressed_config.py
b12x/attention/_shared/mla/compressed_reference.py
b12x/attention/_shared/mla/decode_math.py
b12x/attention/_shared/mla/io.py
b12x/attention/_shared/mla/io_mg.py
b12x/attention/_shared/mla/kernel.py
b12x/attention/_shared/mla/kv_cache.py
b12x/attention/_shared/mla/merge.py
b12x/attention/_shared/mla/packed.py
b12x/attention/_shared/mla/prefill.py
b12x/attention/_shared/mla/prefill_mg.py
b12x/attention/_shared/mla/reference.py
b12x/attention/_shared/mla/smem.py
b12x/attention/_shared/mla/smem_mg.py
b12x/attention/_shared/mla/traits.py
b12x/attention/compressed_mla/__init__.py
b12x/attention/compressed_mla/_scratch.py
b12x/attention/compressed_mla/api.py
b12x/attention/compressed_mla/reference.py
b12x/attention/dense_mla/__init__.py
b12x/attention/dense_mla/_forward.py
b12x/attention/dense_mla/_io.py
b12x/attention/dense_mla/_kernel.py
b12x/attention/dense_mla/_layout.py
b12x/attention/dense_mla/_math.py
b12x/attention/dense_mla/_merge.py
b12x/attention/dense_mla/_reference.py
b12x/attention/dense_mla/_scratch.py
b12x/attention/dense_mla/api.py
b12x/attention/dense_mla/planner.py
b12x/attention/nsa_indexer/__init__.py
b12x/attention/nsa_indexer/_impl.py
b12x/attention/nsa_indexer/api.py
b12x/attention/nsa_indexer/contiguous_kernel.py
b12x/attention/nsa_indexer/fused_indexer.py
b12x/attention/nsa_indexer/kernel.py
b12x/attention/nsa_indexer/msa_reference.py
b12x/attention/nsa_indexer/paged.py
b12x/attention/nsa_indexer/persistent_topk.py
b12x/attention/nsa_indexer/reference.py
b12x/attention/nsa_indexer/schedule_metadata.py
b12x/attention/nsa_indexer/scratch.py
b12x/attention/nsa_indexer/tiled_topk.py
b12x/attention/paged/__init__.py
b12x/attention/paged/_forward.py
b12x/attention/paged/_scratch.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/reference.py
b12x/attention/paged/traits.py
b12x/attention/paged/workspace.py
b12x/attention/paged/tuning/__init__.py
b12x/attention/paged/tuning/registry.py
b12x/attention/sparse_mla/__init__.py
b12x/attention/sparse_mla/_scratch.py
b12x/attention/sparse_mla/api.py
b12x/attention/sparse_mla/reference.py
b12x/attention/varlen/__init__.py
b12x/attention/varlen/api.py
b12x/comm/__init__.py
b12x/comm/pcie/__init__.py
b12x/comm/pcie/_cuda_ipc.py
b12x/comm/pcie/_cute_intrinsics.py
b12x/comm/pcie/_dcp_a2a_cute.py
b12x/comm/pcie/_dcp_cute_common.py
b12x/comm/pcie/_dcp_topk_cute.py
b12x/comm/pcie/_dma_kernels.py
b12x/comm/pcie/_hierarchical_cute.py
b12x/comm/pcie/_launch_geometry.py
b12x/comm/pcie/_oneshot_cute.py
b12x/comm/pcie/_twoshot_cute.py
b12x/comm/pcie/api.py
b12x/comm/pcie/overlap_probe.py
b12x/comm/pcie/pcie_allreduce.py
b12x/comm/pcie/pcie_dcp_a2a.py
b12x/comm/pcie/pcie_dcp_topk.py
b12x/comm/pcie/pcie_dma.py
b12x/comm/pcie/pcie_hierarchical.py
b12x/comm/pcie/pcie_oneshot.py
b12x/comm/pcie/pcie_twoshot.py
b12x/gemm/__init__.py
b12x/gemm/_bmm/__init__.py
b12x/gemm/_bmm/api.py
b12x/gemm/_shared/__init__.py
b12x/gemm/_shared/block_fp8.py
b12x/gemm/_shared/mxfp8_bmm.py
b12x/gemm/_shared/wo_mxfp8.py
b12x/gemm/bf16_gemv/__init__.py
b12x/gemm/bf16_gemv/_kernel.py
b12x/gemm/bf16_gemv/api.py
b12x/gemm/block_fp8_linear/__init__.py
b12x/gemm/block_fp8_linear/api.py
b12x/gemm/blockscaled/__init__.py
b12x/gemm/blockscaled/api.py
b12x/gemm/mla_query_projection/__init__.py
b12x/gemm/mla_query_projection/_bf16.py
b12x/gemm/mla_query_projection/api.py
b12x/gemm/mxfp8_linear/__init__.py
b12x/gemm/mxfp8_linear/_kernel.py
b12x/gemm/mxfp8_linear/api.py
b12x/gemm/tensor_fp8_linear/__init__.py
b12x/gemm/tensor_fp8_linear/_kernel.py
b12x/gemm/tensor_fp8_linear/api.py
b12x/gemm/trellis_linear/__init__.py
b12x/gemm/trellis_linear/_small_m.py
b12x/gemm/trellis_linear/api.py
b12x/gemm/trellis_linear/w4a8.py
b12x/gemm/trellis_linear/csrc/trellis_k6_small.cu
b12x/gemm/trellis_linear/csrc/vendor/LICENSE.exllamav3
b12x/gemm/trellis_linear/csrc/vendor/compat.cuh
b12x/gemm/trellis_linear/csrc/vendor/ptx.cuh
b12x/gemm/trellis_linear/csrc/vendor/util.cuh
b12x/gemm/trellis_linear/csrc/vendor/util.h
b12x/gemm/trellis_linear/csrc/vendor/quant/codebook.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/exl3_devctx.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/exl3_dq.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/exl3_gemm_inner.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/exl3_gemm_kernel.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/exl3_kernel_map.cuh
b12x/gemm/trellis_linear/csrc/vendor/quant/hadamard_inner.cuh
b12x/gemm/wo_projection/__init__.py
b12x/gemm/wo_projection/_quant_cute.py
b12x/gemm/wo_projection/api.py
b12x/integration/__init__.py
b12x/integration/vllm/__init__.py
b12x/integration/vllm/fp6_serving.py
b12x/integration/vllm/plugin.py
b12x/moe/__init__.py
b12x/moe/calibration.py
b12x/moe/_shared/__init__.py
b12x/moe/_shared/execution.py
b12x/moe/_shared/routing.py
b12x/moe/_shared/kernels/__init__.py
b12x/moe/_shared/kernels/activations.py
b12x/moe/_shared/kernels/dynamic.py
b12x/moe/_shared/kernels/micro.py
b12x/moe/_shared/kernels/mxfp6_moe.py
b12x/moe/_shared/kernels/reference.py
b12x/moe/_shared/kernels/reference_flashinfer.py
b12x/moe/_shared/kernels/relu2.py
b12x/moe/_shared/kernels/silu.py
b12x/moe/_shared/kernels/situ.py
b12x/moe/_shared/kernels/tiny_decode.py
b12x/moe/_shared/kernels/trellis_decode.py
b12x/moe/_shared/kernels/trellis_w4a8.py
b12x/moe/_shared/kernels/trellis_w4a8_pair.py
b12x/moe/_shared/kernels/trellis_w4a8_transform.py
b12x/moe/_shared/kernels/trellis_w4a8_uniform.py
b12x/moe/_shared/kernels/w4a8_phase1.py
b12x/moe/_shared/kernels/w4a8_phase2.py
b12x/moe/_shared/kernels/w4a16/__init__.py
b12x/moe/_shared/kernels/w4a16/host.py
b12x/moe/_shared/kernels/w4a16/kernel.py
b12x/moe/_shared/kernels/w4a16/mixed_trellis.py
b12x/moe/_shared/kernels/w4a16/prepare.py
b12x/moe/_shared/kernels/w4a16/route_pack.py
b12x/moe/_shared/kernels/w4a8/__init__.py
b12x/moe/_shared/kernels/w4a8/weights.py
b12x/moe/_shared/kernels/w6a8/__init__.py
b12x/moe/_shared/kernels/w6a8/weights.py
b12x/moe/_shared/tuning/__init__.py
b12x/moe/_shared/tuning/decode_max_active_clusters.py
b12x/moe/_shared/tuning/registry.py
b12x/moe/ep_moe/__init__.py
b12x/moe/ep_moe/_impl.py
b12x/moe/ep_moe/api.py
b12x/moe/fused_moe/__init__.py
b12x/moe/fused_moe/_impl.py
b12x/moe/fused_moe/api.py
b12x/norm/__init__.py
b12x/norm/mhc/__init__.py
b12x/norm/mhc/_impl.py
b12x/norm/mhc/_kernels.py
b12x/norm/mhc/api.py
b12x/quantization/__init__.py
b12x/quantization/mxfp6/__init__.py
b12x/quantization/mxfp6/bf16_to_fp6_small_m.py
b12x/quantization/mxfp6/bf16_to_fp6_tma.py
b12x/quantization/mxfp6/calib_io.py
b12x/quantization/mxfp6/fp6_checkpoint.py
b12x/quantization/mxfp6/fp6_dense_op.py
b12x/quantization/mxfp6/fp6_dense_weights.py
b12x/quantization/mxfp6/fp6_moe_weights.py
b12x/quantization/mxfp6/fp6_row_gs.py
b12x/quantization/mxfp6/fp6_safetensors_export.py
b12x/quantization/mxfp6/fp6_safetensors_load.py
b12x/quantization/mxfp6/model_fp6.py
b12x/quantization/mxfp8/__init__.py
b12x/quantization/mxfp8/api.py
b12x/quantization/nvfp4/__init__.py
b12x/quantization/nvfp4/_impl.py
b12x/quantization/nvfp4/_kernel.py
b12x/quantization/nvfp4/api.py
tests/test_packaging.py
tests/test_registry.py