glm52-spark-kit
GLM-5.2 on 4x DGX Spark (GB10): source-form runtime kit. DCP-aware sparse indexer, fused NVFP4 MLA KV writer, DSpark ring port at DCP4, AEON-base platform image, launchers, gates, session logs. Measured, not projected.
파일 탐색기
- serve_glm52_tp4_dcp4.sh
- tiled_topk.py
- sparse_mla_scratch.py
- b12x_capture.py
- cache.py
- scheduler.py
- vllm.py
- shm_broadcast.py
- parallel_state.py
- arg_utils.py
- builda_bmm_v0.py
- builda_bmm_v1.py
- mla_attention.py
- marlin_moe.py
- logits_processor.py
- sparse_attn_indexer.py
- weight_utils.py
- deepseek_mtp.py
- torch_utils.py
- b12x_mla_sparse.py
- indexer.py
- nvfp4_ds_mla_writer.py
- patch_deep_gemm_ops.py
- sm12x_deep_gemm_fallbacks.py
- sm12x_mqa.py
- registry.py
- __init__.py
- sm12x_mqa.py
- acceptance_length.py
- depth_ladder.py
- kv_cache_interface.py
- PROVENANCE.md
- Dockerfile
- FINDING.md
- LICENSE
- NOTICE
- NVFP4.md
- README.md
- dcp_gather_equivalence.py
- nccl_micro.py
- nccl_p2p_emulation.py
- analyze_trace.py
- profile_drive.sh
- cache_capacity_probe.py
- cachehit_probe.py
- correctness_probe.py
- midctx_probe.py
- patch_gloo_probe.py
- patch_shm_probe.py
- sharedprefix_probe.py
- spec_metrics.py
- probe.py
- DESIGN.md
- floor_bench.sh
- test_capture_hook.py
- test_lane1a_end_to_end.py
- bake_overlays.py
- build_b7.sh
- build_images.sh
- Dockerfile
- Dockerfile.b7
- sidecar_regen.py
- vllm_distributed_parallel_state.py.diff
- vllm_model_executor_layers_attention_mla_attention.py.diff
- vllm_model_executor_layers_sparse_attn_indexer.py.diff
- vllm_utils_torch_utils.py.diff
- vllm_v1_attention_backends_mla_b12x_mla_sparse.py.diff
- vllm_v1_attention_backends_mla_indexer.py.diff
- vllm_v1_kv_cache_interface.py.diff
- vllm_v1_worker_gpu_spec_decode_speculator.py.diff
- content-diff-summary.json
- overlay-diff.json
- 001-index-share-engagement-log.patch
- 002-force-fuse-gemm-comms-env.patch
- 003-launcher-second-rail-and-compaction.patch
- 004-tailquant-router-profiler.patch
- gsm8k-baseline-summary.json
- gsm8k-floor-summary.json
- k5adaptive-partial.jsonl
- prod-clean-final.jsonl
- prod-k2-dcp4-316k.jsonl
- PORTING-NOTES.md
- sircl-gpu-ordering-probe.sh
- tp4_graph_q1_probe.cu
- tp4_probe.cpp
- tp4_tensor_probe.cu
- tp4_vocab_allgather_probe.cu
- tp4_vocab_graph_probe.cu
- transport_probe.cpp
- control_channel.hpp
- eager_staging_timeout.hpp
- gpu_doorbell.hpp
- gpu_graph_command.cuh
- gpu_tp4.hpp
- gpu_tp4_tensor.hpp
- gpu_tp4_vocab_allgather.hpp
- graph_poll_policy.hpp
- memory_buffer.hpp
- statistics.hpp
- topology.hpp
- tp4_allreduce_protocol.hpp
- tp4_c_api.h
- tp4_dual_port_striped_allreduce.hpp
- tp4_graph_command.hpp
- tp4_graph_kernel_strategy.hpp
- tp4_schedule.hpp
- tp4_session.hpp
- tp4_tiled_session.hpp
- tp4_vocab_allgather_c_api.h
- tp4_vocab_allgather_session.hpp
- verbs_endpoint.hpp
- GLM52_TP4_VOCAB_ALLGATHER.md
- probe_shared_capture_stream.py
- probe_vocab_graph_stream_switch.py
- README.md
- sitecustomize.py
- spark_collective_audit.py
- spark_cudagraph_bucket_contract.py
- spark_dcp_collective_audit.py
- spark_graph_status_reporter.py
- spark_persistent_output_ring.py
- spark_tp4_backend.py
- spark_tp4_port_namespace.py
- spark_tp4_query_contract.py
- spark_tp4_query_row_provider.py
- spark_tp4_vocab_allgather_backend.py
- test_backend_install.py
- test_install.py
- test_native_tp4.py
- test_numerical_audit.py
- test_shadow_stats.py
- test_sitecustomize_fail_closed.py
- test_spark_collective_audit.py
- test_spark_cudagraph_bucket_contract.py
- test_spark_dcp_collective_audit.py
- test_spark_graph_status_reporter.py
- test_spark_persistent_output_ring.py
- test_spark_tp4_backend_dispatch.py
- test_spark_tp4_port_namespace.py
- test_spark_tp4_query_contract.py
- test_spark_tp4_query_row_provider.py
- test_spark_tp4_vocab_allgather_backend.py
- test_vocab_graph_stream_switch_regression.py
- tp4_numerical_audit.py
- nccl-2.29.7-skip-tree-pat.patch
- nccl-2.30.7-advertise-all-listener-gids.patch
- nccl-2.30.7-skip-tree-pat.patch
- probe_dcp2_collectives.py
- probe_dcp4_collectives.py
- probe_dcp4_entrypoint.sh
- probe_entrypoint.sh
- README.md
- qualify_direct_cable.py
- run_tp4_graph_q1_probe.ps1
- run_tp4_numerical_audit.ps1
- run_tp4_probe.ps1
- run_tp4_tensor_probe.ps1
- run_tp4_vocab_allgather_probe.ps1
- run_tp4_vocab_graph_probe.ps1
- run_tp4_vocab_graph_stream_switch_probe.ps1
- test_qualify_direct_cable.py
- control_channel.cpp
- cuda_event_gate.hpp
- cuda_stream_handoff.hpp
- gpu_doorbell.cu
- gpu_tp4.cu
- gpu_tp4_tensor.cu
- gpu_tp4_vocab_allgather.cu
- memory_buffer.cu
- statistics.cpp
- topology.cpp
- tp4_c_api.cpp
- tp4_dual_port_striped_host.hpp
- tp4_graph_command.cpp
- tp4_schedule.cpp
- tp4_session.cpp
- tp4_vocab_allgather_c_api.cpp
- tp4_vocab_allgather_session.cpp
- verbs_endpoint.cpp
- verbs.h
- cuda_runtime.h
- cuda_event_gate_test.cu
- eager_staging_timeout_test.cpp
- gpu_doorbell_layout_test.cpp
- graph_poll_policy_test.cpp
- statistics_test.cpp
- test_topology_schedule_source_contract.py
- test_tp4_deferred_ack_source_contract.py
- test_tp4_dual_port_striped_gpu_source_contract.py
- test_tp4_dual_port_striped_host_source_contract.py
- test_tp4_graph_prefill_probe_contract.py
- test_tp4_split_graph_source_contract.py
- topology_test.cpp
- tp4_allreduce_protocol_test.cpp
- tp4_c_api_layout_test.cpp
- tp4_c_api_test.cpp
- tp4_dual_port_striped_allreduce_test.cpp
- tp4_dual_port_striped_host_test.cpp
- tp4_graph_command_test.cpp
- tp4_graph_kernel_strategy_test.cpp
- tp4_schedule_test.cpp
- tp4_vocab_allgather_c_api_test.cpp
- tp4_vocab_allgather_layout_test.cpp
- verbs_completion_test.cpp
- wire_protocol_test.cpp
- CABLE_QUALIFICATION.md
- CMakeLists.txt
- README.md
- LICENSE
- NOTICE
- THIRD_PARTY_NOTICES.md
- UPSTREAM.md
- encode_vec2.py
- mcg_codec.py
- roundtrip.py
- btx-layer-00000.safetensors
- btx-manifest.json
- btx-layer-00000.safetensors
- btx-manifest.json
- exl3_comp_unit_1.cuh
- exl3_comp_unit_1_cb0.cu
- exl3_comp_unit_1_cb1.cu
- exl3_comp_unit_1_cb2.cu
- exl3_comp_unit_2.cuh
- exl3_comp_unit_2_cb0.cu
- exl3_comp_unit_2_cb1.cu
- exl3_comp_unit_2_cb2.cu
- exl3_comp_unit_3.cuh
- exl3_comp_unit_3_cb0.cu
- exl3_comp_unit_3_cb1.cu
- exl3_comp_unit_3_cb2.cu
- exl3_comp_unit_4.cuh
- exl3_comp_unit_4_cb0.cu
- exl3_comp_unit_4_cb1.cu
- exl3_comp_unit_4_cb2.cu
- exl3_comp_unit_5.cuh
- exl3_comp_unit_5_cb0.cu
- exl3_comp_unit_5_cb1.cu
- exl3_comp_unit_5_cb2.cu
- exl3_comp_unit_6.cuh
- exl3_comp_unit_6_cb0.cu
- exl3_comp_unit_6_cb1.cu
- exl3_comp_unit_6_cb2.cu
- exl3_comp_unit_7.cuh
- exl3_comp_unit_7_cb0.cu
- exl3_comp_unit_7_cb1.cu
- exl3_comp_unit_7_cb2.cu
- exl3_comp_unit_8.cuh
- exl3_comp_unit_8_cb0.cu
- exl3_comp_unit_8_cb1.cu
- exl3_comp_unit_8_cb2.cu
- exl3_gemv_int8_inst_coop_k1.cu
- exl3_gemv_int8_inst_coop_k2.cu
- exl3_gemv_int8_inst_coop_k3.cu
- exl3_gemv_int8_inst_coop_k4.cu
- exl3_gemv_int8_inst_coop_k5.cu
- exl3_gemv_int8_inst_coop_k6.cu
- exl3_gemv_int8_inst_coop_k7.cu
- exl3_gemv_int8_inst_coop_k8.cu
- exl3_gemv_int8_inst_sq_k1.cu
- exl3_gemv_int8_inst_sq_k2.cu
- exl3_gemv_int8_inst_sq_k3.cu
- exl3_gemv_int8_inst_sq_k4.cu
- exl3_gemv_int8_inst_sq_k5.cu
- exl3_gemv_int8_inst_sq_k6.cu
- exl3_gemv_int8_instances.cuh
- exl3_moe_inst_k0_n128_cb1.cu
- exl3_moe_inst_k0_n128_cb2.cu
- exl3_moe_inst_k0_n256_cb1.cu
- exl3_moe_inst_k0_n256_cb2.cu
- exl3_moe_inst_k1_cb1.cu
- exl3_moe_inst_k1_cb2.cu
- exl3_moe_inst_k2_cb1.cu
- exl3_moe_inst_k2_cb2.cu
- exl3_moe_inst_k3_cb1.cu
- exl3_moe_inst_k3_cb2.cu
- exl3_moe_inst_k4_cb1.cu
- exl3_moe_inst_k4_cb2.cu
- exl3_moe_inst_k5_cb1.cu
- exl3_moe_inst_k5_cb2.cu
- exl3_moe_inst_k6_cb1.cu
- exl3_moe_inst_k6_cb2.cu
- exl3_moe_inst_k7_cb1.cu
- exl3_moe_inst_k7_cb2.cu
- exl3_moe_inst_k8_cb1.cu
- exl3_moe_inst_k8_cb2.cu
- exl3_moe_instances.cuh
- quantize_tiles_inst_k1.cu
- quantize_tiles_inst_k2.cu
- quantize_tiles_inst_k3.cu
- quantize_tiles_inst_k4.cu
- quantize_tiles_inst_k5.cu
- quantize_tiles_inst_k6.cu
- quantize_tiles_inst_k7.cu
- quantize_tiles_inst_k8.cu
- quantize_tiles_instances.cuh
- codebook.cuh
- coop_autotune.cu
- coop_autotune.cuh
- exl3_devctx.cu
- exl3_devctx.cuh
- exl3_dq.cuh
- exl3_gemm.cu
- exl3_gemm.cuh
- exl3_gemm_inner.cuh
- exl3_gemm_kernel.cuh
- exl3_gemv.cu
- exl3_gemv.cuh
- exl3_gemv_int8.cu
- exl3_gemv_int8.cuh
- exl3_gemv_int8_kernel.cuh
- exl3_gemv_kernel.cuh
- exl3_kernel_map.cu
- exl3_kernel_map.cuh
- exl3_moe.cu
- exl3_moe.cuh
- exl3_moe_common.cuh
- exl3_moe_kernel.cuh
- hadamard.cu
- hadamard.cuh
- hadamard_inner.cuh
- pack.cu
- pack.cuh
- quantize.cu
- quantize.cuh
- quantize_tiles_kernel.cuh
- reconstruct.cu
- reconstruct.cuh
- util.cu
- util.cuh
- __init__.py
- quantize.py
- __init__.py
- exl3.py
- fp16.py
- LICENSE
- UPSTREAM.md
- btx_writer.py
- DESIGN-v3.md
- DESIGN.md
- gen_uniform_k2_oracle_inputs.py
- oracle_result_20260825.json
- probe_btx_mixed.py
- probe_btx_mixed_toy.py
- run_uniform_k2_oracle.sh
- split.py
- ASK-RESULTS-20260824.md
- CAMPAIGN-STATE.md
- D2-VERDICT-20260825.md
- d3-transport-comparison.md
- JOURNAL.md
- LANE3-PORT-MAP.md
- PHASE3-PLAN.md
- README.md
- run_phase2.sh
- velogb10-transport-intel.md
- win3_identity_diff.sh
- win3_replay_200.jsonl
- win3_replay_set.py
- win3_replay_stream.py
- win3_sampler.py
- WINDOW-1-MANIFEST.md
- WINDOW-2-MANIFEST.md
- WINDOW-3-MANIFEST.md
- WINDOW-K-CLOSEOUT.md
- CAMPAIGN-2026-08.md
- COMM_FLOOR-2026-08-19.md
- DECODE_STEP_PROFILE-2026-08-18.md
- DSPARK.md
- HANDOFF-2026-08-18.md
- NVFP4_DS_MLA_RECORD.md
- RECOMMENDATION.md
- SESSION-2026-08-17.md
- TAILQUANT-PROFILER.md
- dspark-ddp.sh
- dspark_finetune.py
- glmdatadrive.py
- glmdeepdrive.py
- PROVENANCE.md
- nvfp4_writer_gate.py
- vllm-b12x.Dockerfile
- marlin_w8a16_baseline.py
- README.md
- w8a16_smallm_gemv.py
- w8a16_smallm_gemv_v4.py
- ag4_proto.c
- README.md
- nvfp4_ds_mla_writer.py
- nvfp4_writer_triton.standalone.py
- champion_adaptive.sh
- champion_dspark_ring_32k.sh
- champion_dspark_ring_dcp4_32k.sh
- champion_fusedgather_32k.sh
- champion_indexer.sh
- dspark_ddp_finetune.sh
- launch_gx10.sh
- node_protect.sh
- nvfp4_swap.sh
- rails_netplan.sh
- resolve_gid_and_launch.sh
- wedge_capture.sh
- tiled_topk.py
- sparse_mla_scratch.py
- cp_utils.py
- dflash_speculator.py
- kv_cache_manager.py
- qwen3_dflash.py
- speculators_algos.py
- README.md
- sparse_attn_indexer.stock+fused.py
- marlin_moe_profiler.py
- b12x_capture.py
- cache.py
- scheduler.py
- speculative.py
- vllm.py
- shm_broadcast.py
- parallel_state.py
- arg_utils.py
- builda_bmm_v0.py
- builda_bmm_v1.py
- mla_attention.py
- marlin_moe.py
- logits_processor.py
- sparse_attn_indexer.py
- weight_utils.py
- deepseek_mtp.py
- torch_utils.py
- b12x_mla_sparse.py
- indexer.py
- nvfp4_ds_mla_writer.py
- patch_deep_gemm_ops.py
- sm12x_deep_gemm_fallbacks.py
- sm12x_mqa.py
- sparse_utils.py
- registry.py
- __init__.py
- sm12x_mqa.py
- scheduler.py
- acceptance_length.py
- depth_ladder.py
- speculator.py
- cudagraph_utils.py
- kv_cache_interface.py
- 0001-speculators-dspark-VLLM_DSPARK_DRAFT_WINDOW-opt-in-w.patch
- 0002-dspark-dflash-ring-buffer-draft-KV-VLLM_DSPARK_DRAFT.patch
- 0003-dspark-ring-deterministic-ingest-keep-only-trailing-.patch
- 0004-cp_utils-exempt-spec-less-layers-from-DCP-attention-.patch
- 0005-dspark-ring-kill-per-step-host-sync-in-bookkeeping.patch
- 0006-dspark-ring-sync-free-ingest-via-trash-slot-redirect.patch
- 0007-weight_utils-tolerate-non-dict-hf_overrides-in-get_q.patch
- 0008-dflash-ingest-quant-aware-fallback-for-packed-draft-.patch
- 0009-dflash-attention-quant-aware-qkv-projection-in-draft.patch
- 0010-dflash-ingest-fuse-dequantized-KV-weights-for-quanti.patch
- 0011-dspark-ring-skip-mask-rebuild-in-steady-long-context.patch
- 0012-dspark-env-gated-training-data-capture-VLLM_DSPARK_C.patch
- 0013-dspark-capture-v2-raw-per-layer-aux-hiddens-configur.patch
- 0014-dspark-capture-token-ids-from-the-target-batch-not-d.patch
- 0015-dspark-ring-trim-prefix-cache-hits-by-the-draft-wind.patch
- 0016-dspark-confidence-gated-variable-draft-submission-bs.patch
- 0017-dspark-ring-hoist-slot-addressing-out-of-the-per-lay.patch
- 0020-dspark-capture-payload-builder.patch
- 0021-dspark-capture-min-t-guard.patch
- 0090-cp_utils-exempt-spec-less-layers-from-DCP-attention-.patch
- 0091-kv_cache-host-non-MLA-draft-layers-in-the-full-MLA-g.patch
- 0092-weight_utils-tolerate-non-dict-hf_overrides-in-get_q.patch
- PORT-NOTES.md
- kv_cache_interface.py
- torch_utils.py
- Dockerfile
- README.md
- CAMPAIGN-BATTERY-2026-08-24.md
- b12x_sparse_helpers.py
- deepseek_v2.py
- flashmla_sparse.py
- NOTICE
- patch_flashmla_ops.py
- sm12x_deep_gemm_fallbacks.py
- sm12x_mqa.py
- sm12x_sparse_mla_attn.py
- sparse_attn_indexer.py
- sparse_mla_env.py
- sparse_mla_kernels.py
- apply_scheduler_patch.py
- config_args.patch
- README.md
- run.sh
- run.sh
- run.sh
- build-nvfp4-gx10.sh
- README.md
- fix-indexer-mtp-overhang.py
- Dockerfile
- Dockerfile.v2
- Dockerfile.v2b
- draft-quant.patch
- launch_gx10.sh
- launch_gx10_da2048.sh
- launch_gx10_da256.sh
- launch_gx10_fp8peak.sh
- launch_gx10_k4.sh
- launch_gx10_nospec.sh
- README.md
- resolve_gid_and_launch.sh
- fix-indexer-mtp-overhang.py
- fix-scheduler-preempt-race.py
- launcher-rank-verification.sh.excerpt
- glm-dcp2-v4-speed128k-adaptive.yaml
- INDEXER-DCP-MTP-BUG.md
- README.md
- keyerror_exact_error_text.txt
- README.md
- README.md
- AGENT-BRIEF.md
- Dockerfile.deepgemm-fix
- Dockerfile.fix2
- install_deepgemm.sh
- launch_gx10_v027.sh
- MEMORY-AND-KERNEL-FINDINGS.md
- patch_masked_mha.py
- README.md
- screen027.jsonl
- screen_027.sh
- SCREENING.md
- wedge_trigger.py
- cell0.json
- cellA-nvfp4.jsonl
- cellA2b.jsonl
- cellCapfit.jsonl
- cellK2.jsonl
- challenger-v2.json
- challenger-v2b.json
- challenger-v2c.json
- challenger.json
- kernel-exp1.json
- kernel-exp2.json
- probe_battery.py
- LICENSE
- NOTICE
- README.md
- test_dcp_topk_pack_identity.py
- test_nvfp4_writer_identity.py
- test_nvfp4_writer_opcheck.py
- .gitignore
- apply.sh
- LICENSE
- MANIFEST.json
- NOTICE
- README.md
- THIRD_PARTY_NOTICES.md
- verify.sh
# CDN으로 사용하기
jsDelivrjsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.
링크
예시
// repository documentation
Was this content helpful?
(0 ratings)