{ "schema": "kbp-card/v1", "blueprint_id": "preview:b707d03d56bb6c007939c0cceb452f91a17a938512f22ef1dd7a82a3c608bda8", "lineage": "sha256:2ed80c3cc566ea7b2e0b5cc1763fe6c6e9f6bbafe59c2d83e9dd706fe9369125", "revision": 0, "name": "glm-5.3-flash-nvfp4-tp3-512k", "slug": "glm-5.3-flash-nvfp4-tp3-512k--gb10", "preview": true, "exported_by": "preview: tools/preview/kbp_preview.py over `met circuit plan` (not `met blueprint export`)", "model": { "repo": "nvidia/GLM-5.3-Flash-NVFP4", "revision": null, "arch": "glm5_next", "params": "320B-A18B", "dtype": "nvfp4", "quantization": "NVFP4 experts, W4A16 dense, W8A8 DSA projections", "layers": 45, "layer_kinds": { "linear_attention": 34, "deepseek_sparse_attention": 11 }, "max_seq_len": 524288, "description": "GLM-5 hybrid: KDA and DSA sparse latent attention under 4-stream hyper-connections, dense SwiGLU then 288-expert MoE, MTP draft head" }, "hardware": { "class": "gb10", "device": "gb10", "name": "NVIDIA GB10 (DGX Spark)", "ptx_arch": "sm_121f", "sm_count": 48, "memory_gb": 128, "memory_bandwidth_gbps": 273, "world_size": 3, "min_visible_mib": null, "devices": [ "gb10" ] }, "precision": { "tier": "recipe", "table": "glm-5.3-flash-nvfp4-tp3", "table_sha256": "f1e299043e34000299680b8d2fce60b62342ccc487d9cb9734c9ddc7dca6d15a", "linear_by_format": [ { "weight": "nvfp4/g16", "activation": "bf16", "nodes": 322 }, { "weight": "fp8/channel", "activation": "fp8/token", "nodes": 123 }, { "weight": "bf16", "activation": "f32", "nodes": 90 }, { "weight": "nvfp4/g16", "activation": "nvfp4/g16", "nodes": 90 }, { "weight": "bf16", "activation": "bf16", "nodes": 87 } ], "outputs_by_format": [ { "format": "bf16", "nodes": 904 }, { "format": "f32", "nodes": 411 }, { "format": "fp8/token", "nodes": 101 }, { "format": "bf16,f32,f32", "nodes": 90 }, { "format": "nvfp4/g16", "nodes": 90 }, { "format": "f32,i32", "nodes": 42 }, { "format": "f32,f32", "nodes": 34 }, { "format": "i32", "nodes": 12 } ], "kv_cache": "fp8" }, "execution": "legacy:glm5_next", "kernels": { "count": 22, "entries": 34, "families": 20, "residual": 0, "launches": 1411, "groups": { "device": 1213, "host": 2, "novel": 84 }, "compute": { "cuda_core": 505, "memory": 506, "tensor_core:mma.sync.m16n8k32.e4m3": 112, "tensor_core:mma.sync.m16n8k64.mxf4nvf4.block_scale": 90 }, "kinds": null, "bytes_raw": null, "bytes_deflated": null, "sources": null, "list": [ { "id": "causal_conv1d::causal_conv1d_update_l2norm", "family": "causal_conv1d_l2norm", "groups": 34, "compute": [ "cuda_core" ] }, { "id": "dsa_indexer::dsa_indexer_store", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "dsa_indexer_rows::dsa_expand_selection_rows", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "dsa_scores_warp::dsa_index_scores_heads_paged", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "dsa_scores_warp::dsa_pool_keys_store_rows", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "dsa_topk_mc::dsa_topk_mc_collect", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "dsa_topk_mc::dsa_topk_mc_pass", "family": "dsa_indexer", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "gemv::dense_gemv_bf16", "family": "dense_bf16", "groups": 12, "compute": [ "cuda_core" ] }, { "id": "gemv::dense_gemv_bf16_fp32out", "family": "dense_bf16", "groups": 64, "compute": [ "cuda_core" ] }, { "id": "glm5next_dsa_mla_decode::glm5next_dsa_mla_decode_fp8", "family": "glm_mla_decode", "groups": 11, "compute": [ "cuda_core" ] }, { "id": "glm5next_ffn::glm5next_moe_combine", "family": "glm_moe_combine", "groups": 42, "compute": [ "memory" ] }, { "id": "glm5next_ffn::glm5next_router_topk_rank", "family": "glm_router_topk", "groups": 42, "compute": [ "memory" ] }, { "id": "glm5next_ffn::glm5next_swiglu_clamp", "family": "glm_swiglu_clamp", "groups": 87, "compute": [ "memory" ] }, { "id": "glm5next_mhc::glm5next_hc_expand", "family": "glm_mhc", "groups": 1, "compute": [ "memory" ] }, { "id": "glm5next_mhc::glm5next_hc_head", "family": "glm_mhc", "groups": 1, "compute": [ "memory" ] }, { "id": "glm5next_mhc::glm5next_hc_post", "family": "glm_mhc", "groups": 90, "compute": [ "memory" ] }, { "id": "glm5next_mhc_c1::glm5next_hc_pre_norm_c1", "family": "glm_mhc", "groups": 90, "compute": [ "cuda_core" ] }, { "id": "glm5next_mla_latent_write::glm5next_mla_latent_write_fp8", "family": "glm_mla_latent_write", "groups": 11, "compute": [ "memory" ] }, { "id": "kda_gate::kda_gate_bf16", "family": "kda_gates", "groups": 34, "compute": [ "memory" ] }, { "id": "kda_layer_ops::kda_o_norm_gated_bf16", "family": "kda_o_norm", "groups": 34, "compute": [ "memory" ] }, { "id": "kda_layer_ops::kda_pack_qkv_bf16", "family": "kda_pack", "groups": 34, "compute": [ "memory" ] }, { "id": "kda_layer_ops::kda_sigmoid_bf16_f32", "family": "kda_gates", "groups": 34, "compute": [ "memory" ] }, { "id": "kda_recurrent::kda_recurrent_decode_bf16_smem", "family": "kda_recurrent", "groups": 34, "compute": [ "cuda_core" ] }, { "id": "nllb_encoder::nllb_layernorm_bf16", "family": "layer_norm", "groups": 11, "compute": [ "memory" ] }, { "id": "norm::rms_norm", "family": "rms_norm", "groups": 1, "compute": [ "memory" ] }, { "id": "rms_norm_vanilla::rms_norm_vanilla", "family": "rms_norm", "groups": 11, "compute": [ "memory" ] }, { "id": "w4a16_gemv::w4a16_gemv", "family": "w4a16_gemv", "groups": 238, "compute": [ "cuda_core" ] }, { "id": "w4a4_gemv_mx::w4a4_gemv_mx8", "family": "w4a4_mx", "groups": 6, "compute": [ "tensor_core:mma.sync.m16n8k64.mxf4nvf4.block_scale" ] }, { "id": "w4a4_gemv_mx::w4a4_quant_rows", "family": "w4a4_mx", "groups": 6, "compute": [ "memory" ] }, { "id": "w4a4_gemv_mx_moe::w4a4_gemv_mx8_moe_slots_sweep_cs", "family": "w4a4_mx_static", "groups": 84, "compute": [ "tensor_core:mma.sync.m16n8k64.mxf4nvf4.block_scale" ] }, { "id": "w4a4_gemv_mx_moe::w4a4_quant_rows_static", "family": "w4a4_mx_static", "groups": 84, "compute": [ "tensor_core:mma.sync.m16n8k64.mxf4nvf4.block_scale" ] }, { "id": "w8a8_act_quant::w8a8_act_quant_row", "family": "w8a8_act_quant", "groups": 101, "compute": [ "memory" ] }, { "id": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8", "family": "wxay", "groups": 101, "compute": [ "tensor_core:mma.sync.m16n8k32.e4m3" ] }, { "id": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8_split4", "family": "wxay", "groups": 11, "compute": [ "tensor_core:mma.sync.m16n8k32.e4m3" ] } ] }, "plans": [ { "mode": "decode", "rows": 1, "digest": "1ff131e588e38681729df0cb2ae65a4d9adc8cf91eeb184a1b8bfb7527ccb5fb", "groups": 1299, "launches": 1411, "copies": 0, "edges_fused": 396, "edges_materialized": 1544, "routes": [] }, { "mode": "multi_seq", "rows": 2, "digest": "6901b4656acaaf7d12f09ccd77ab60bcaa1696dd7ebdcbb53b5f07bf99e2729f", "groups": 1162, "launches": 1790, "copies": 0, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "multi_seq", "rows": 4, "digest": "97f9f6bb5b992cf3bd2a42dbdea856f941905452ad466044498902d8d6a10d9f", "groups": 1162, "launches": 2212, "copies": 0, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "multi_seq", "rows": 8, "digest": "b7345097c70353d20e2ee062e090ed5993bc9bff45011cf32ad8726d7c13fdca", "groups": 1162, "launches": 3056, "copies": 0, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "multi_seq", "rows": 12, "digest": "c30a839202ebafdcd5ce7478c5116ac40297816ce59385f79a0fd04cfd6d59cf", "groups": 1184, "launches": 3922, "copies": 0, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "multi_seq", "rows": 16, "digest": "c39ed89d8b0fabba3184d6af97f6d1ed71fa74959bbfae24b3121ae5ac9cf13e", "groups": 1184, "launches": 4766, "copies": 0, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "multi_seq", "rows": 24, "digest": "d242da1f6372bec8b7e0b36c0a46ed2d87d6fae1be2b65de557c1adf04a93e23", "groups": 1672, "launches": 295, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "f2d841d37452974cc9d17e093ff63bfc7f4010646943be695e025bf5690558bc", "groups": 1672, "launches": 1077, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "multi_seq", "rows": 32, "digest": "415a2b942fc2f0b326602c22e2fc36a1a385f949ea16490fb6d3f4a554134204", "groups": 1672, "launches": 295, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "501954598eb52fb2b23cf5d7b71c0c411d8517c855aae0ecf1b7e4450874ac3e", "groups": 1672, "launches": 1349, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "multi_seq", "rows": 48, "digest": "5ed8f0301f4e79e2a8dec20b6187fd37d9810cf563b4ca621466904d6264501b", "groups": 1672, "launches": 196, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "0466e915dcab4a3a2e39714d199b22a9360e7e66f82fdd8df649568d47c8ab36", "groups": 1672, "launches": 1794, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "multi_seq", "rows": 64, "digest": "ae9c63121d471006dab5fb16a6633c4c0a934d7534a8f9dc50abf519d7c6b228", "groups": 1672, "launches": 196, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "b1a28ab12a545a55a803d579cc2c67ce93afd950628053d45d7495422a42c827", "groups": 1672, "launches": 2338, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "multi_seq", "rows": 96, "digest": "e908fdfdccb6266d78c82148748a88673c46f83903903ba569026906c32c83fb", "groups": 1672, "launches": 196, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "e90a27d394bfd3af74bf400fde78eac35b4de42f051c69e9cac7153a5e515a03", "groups": 1672, "launches": 3426, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "multi_seq", "rows": 128, "digest": "2a6b51a59e1f248d2972925af3219f37b5d5948d91b91c73371a9884c350626f", "groups": 1672, "launches": 196, "copies": 0, "edges_fused": 34, "edges_materialized": 1906, "routes": [ { "route": "gdn_state_slots_fragmented", "when": "when the sequences' GDN state slots are not contiguous in slice order (slots fragment as sequences finish), and the step runs the per-sequence loop; planned as `ssm_batched_recurrent=off` (ml/qwen3_ssm/trait_decode_multi_seq/ssm_batched_recurrent.rs:93-150 (the contiguity check), :396-397 (the per-sequence call))", "digest": "30d2adaf14824cdc55d76f358868856c5ba6b166be1d89ee3f294a7dc9c08ab1", "groups": 1672, "launches": 4514, "copies": 0, "edges_fused": 34, "edges_materialized": 1906 } ] }, { "mode": "verify", "rows": 2, "digest": "5d2226cbcbc189e76766595702a1c403eadfb404408b147fa0215df818afa635", "groups": 1173, "launches": 1836, "copies": 34, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "verify", "rows": 3, "digest": "3446f07c8c410b4e32cb9d065f8a6f2afa111cf9dedab48b6fd0d836be4fa4c6", "groups": 1173, "launches": 2081, "copies": 68, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "verify", "rows": 4, "digest": "063961fa91fb7131b3bdd9ab41e9c8236a4307f474136c57f95a931fd836f38d", "groups": 1173, "launches": 2326, "copies": 102, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "verify", "rows": 8, "digest": "3bc1c08e84d29eb84beb6e53b5dee5b74f195388ba6be44e394a0c66131a7f38", "groups": 1173, "launches": 3298, "copies": 238, "edges_fused": 522, "edges_materialized": 1418, "routes": [] }, { "mode": "draft", "rows": 1, "digest": "8969b4c107e71640fdf1a199dcdc0ddc592ee3de57c33ba29a2fbe413a36ca62", "groups": 37, "launches": 12, "copies": 0, "edges_fused": 0, "edges_materialized": 38, "routes": [] } ], "sizes": { "body": null, "lite_file": null, "full_file": null }, "signature": { "status": "unverified", "suite": "mldsa65+ed25519", "kid": null, "anchor": null, "verified_at": null }, "download": { "available": false, "note": "No .kbp file exists for a preview: blueprint builds and signing are not available yet." }, "performance": [ { "benchmark": "ctxsweep-c1", "metric": "decode_tok_s", "value": 50.52, "unit": "tok/s", "box": "gb10", "label": "C=1 decode, 1k context", "context": "1k", "samples": [ 50.12, 50.52, 50.8 ], "status": "preliminary", "note": "median of runs p1-p3 (recipe defaults); single harness, bring-up ctxsweep, C=1, OSL 512", "source": { "file": "profile glm53-tp3-512k-fast12/PROVENANCE.txt", "lines": null, "commit": null, "record": "glm53-tp3-512k-fast12: fast11 recipe verbatim; binary sha256 377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8 = wt-dsactx 6d845c0ee (95fccbf63 + METRALE_LISTENER_RCVBUF kill switch + M" }, "fingerprint": { "record_sha256": "9e928d6097467ec8380cc92e7f8a9ab26c54bb04f439ea6065c786a7b93e406b", "closure": null, "argv_sha256": "e17fe83b2704457eb7d591072bc901742e581c2ff5863602bd048f21dfe0f70c", "env_sha256": "07768e9f8fd78df2c7f624752990c30a98c8c01497e58bbe28d301f60179f557", "draw_sha256": null, "git_sha": "6d845c0ee", "binary_sha256": "377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8" } }, { "benchmark": "ctxsweep-c1", "metric": "decode_tok_s", "value": 49.43, "unit": "tok/s", "box": "gb10", "label": "C=1 decode, 16k context", "context": "16k", "samples": [ 49.07, 49.43, 49.52 ], "status": "preliminary", "note": "median of runs p1-p3 (recipe defaults); single harness, bring-up ctxsweep, C=1, OSL 512", "source": { "file": "profile glm53-tp3-512k-fast12/PROVENANCE.txt", "lines": null, "commit": null, "record": "glm53-tp3-512k-fast12: fast11 recipe verbatim; binary sha256 377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8 = wt-dsactx 6d845c0ee (95fccbf63 + METRALE_LISTENER_RCVBUF kill switch + M" }, "fingerprint": { "record_sha256": "9e928d6097467ec8380cc92e7f8a9ab26c54bb04f439ea6065c786a7b93e406b", "closure": null, "argv_sha256": "e17fe83b2704457eb7d591072bc901742e581c2ff5863602bd048f21dfe0f70c", "env_sha256": "07768e9f8fd78df2c7f624752990c30a98c8c01497e58bbe28d301f60179f557", "draw_sha256": null, "git_sha": "6d845c0ee", "binary_sha256": "377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8" } }, { "benchmark": "ctxsweep-c1", "metric": "decode_tok_s", "value": 47.96, "unit": "tok/s", "box": "gb10", "label": "C=1 decode, 64k context", "context": "64k", "samples": [ 47.7, 47.96, 48.11 ], "status": "preliminary", "note": "median of runs p1-p3 (recipe defaults); single harness, bring-up ctxsweep, C=1, OSL 512", "source": { "file": "profile glm53-tp3-512k-fast12/PROVENANCE.txt", "lines": null, "commit": null, "record": "glm53-tp3-512k-fast12: fast11 recipe verbatim; binary sha256 377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8 = wt-dsactx 6d845c0ee (95fccbf63 + METRALE_LISTENER_RCVBUF kill switch + M" }, "fingerprint": { "record_sha256": "9e928d6097467ec8380cc92e7f8a9ab26c54bb04f439ea6065c786a7b93e406b", "closure": null, "argv_sha256": "e17fe83b2704457eb7d591072bc901742e581c2ff5863602bd048f21dfe0f70c", "env_sha256": "07768e9f8fd78df2c7f624752990c30a98c8c01497e58bbe28d301f60179f557", "draw_sha256": null, "git_sha": "6d845c0ee", "binary_sha256": "377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8" } }, { "benchmark": "ctxsweep-c1", "metric": "decode_tok_s", "value": 47.96, "unit": "tok/s", "box": "gb10", "label": "C=1 decode, 128k context", "context": "128k", "samples": [ 47.67, 47.96, 47.98 ], "status": "preliminary", "note": "median of runs p1-p3 (recipe defaults); single harness, bring-up ctxsweep, C=1, OSL 512; runs p2 and p3 at 128k are throttle-flagged in the record", "source": { "file": "profile glm53-tp3-512k-fast12/PROVENANCE.txt", "lines": null, "commit": null, "record": "glm53-tp3-512k-fast12: fast11 recipe verbatim; binary sha256 377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8 = wt-dsactx 6d845c0ee (95fccbf63 + METRALE_LISTENER_RCVBUF kill switch + M" }, "fingerprint": { "record_sha256": "9e928d6097467ec8380cc92e7f8a9ab26c54bb04f439ea6065c786a7b93e406b", "closure": null, "argv_sha256": "e17fe83b2704457eb7d591072bc901742e581c2ff5863602bd048f21dfe0f70c", "env_sha256": "07768e9f8fd78df2c7f624752990c30a98c8c01497e58bbe28d301f60179f557", "draw_sha256": null, "git_sha": "6d845c0ee", "binary_sha256": "377d5214f286c6506841271a8de6a646ce9900a152c4327fbe9f2864b9f255e8" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 191.072, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "shared expert down M=2 N=4096 K=768", "floor_us": 7.184, "pct_of_floor": 3.76, "bound": "memory", "status": "measured", "note": "shared expert down M=2 N=4096 K=768; n=10343 calls, 14.4% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; side stream, co-runs with the routed W4A4 sweep; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3001, 3014 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "8feeba49fb41ab8c38f68a868fa40c236d9a9c024f3baf7ef5e349ed239303b4", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 103.328, "unit": "us", "box": "gb10", "label": "glm5next_moe_combine", "kernel": "glm5next_ffn::glm5next_moe_combine", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "rows=2 top_k=8 H=4096", "floor_us": 0.658, "pct_of_floor": 0.64, "bound": "memory", "status": "measured", "note": "rows=2 top_k=8 H=4096; n=10343 calls, 9.3% of regime kernel time; peak class bf16; routed slots + shared in, BF16 out; median solo 1.44 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3016, 3029 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "82b48821e8d5a9e9fa5270aa5cc7d410d002f49e567ccb4af873ac0d0fc36b75", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 90.368, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096", "floor_us": 80.948, "pct_of_floor": 89.58, "bound": "memory", "status": "measured", "note": "KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096; n=8375 calls, 6.6% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; median solo 85.44 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3031, 3044 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "f8ff990c4ad3d8b388566aa6db294d370a9458f855b7efde6fdd91730f898fe1", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 201.984, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "DSA o_absorb M=2 N=4096 K=11264", "floor_us": 185.513, "pct_of_floor": 91.85, "bound": "memory", "status": "measured", "note": "DSA o_absorb M=2 N=4096 K=11264; n=2709 calls, 4.8% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 199.52 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3046, 3059 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "065db9f9b9b1e8d1a4b88a61a1c0d332b83c74f10d6bc04acca6264a1c4faec1", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 34.624, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "KDA o_proj M=2 N=4096 K=2816", "floor_us": 26.168, "pct_of_floor": 75.58, "bound": "memory", "status": "measured", "note": "KDA o_proj M=2 N=4096 K=2816; n=8375 calls, 2.5% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; median solo 33.79 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3061, 3074 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "51e5fc3100bd2d6ce1daf8c84484160c7cfb2e63e8fcdd7a7397b23e311f9815", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 25.216, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "shared expert gate|up M=2 N=768+768 K=4096", "floor_us": 14.303, "pct_of_floor": 56.72, "bound": "memory", "status": "measured", "note": "shared expert gate|up M=2 N=768+768 K=4096; n=10343 calls, 2.3% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; side stream; median solo 12.48 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3076, 3089 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "d52b61a0bbae51a30197ab2e6e078386a69c460a410f4b9d79581d197539dbef", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 10.624, "unit": "us", "box": "gb10", "label": "glm5next_hc_pre_norm_c1", "kernel": "glm5next_mhc_c1::glm5next_hc_pre_norm_c1", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 3.851, "pct_of_floor": 36.24, "bound": "memory", "status": "measured", "note": "T=2 hc=4 H=4096 mix=24; n=22168 calls, 2.2% of regime kernel time; peak class bf16; BF16 hc_fn x f32 streams on CUDA cores; median solo 10.62 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3091, 3104 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "8fa57ad78499b5314f9c7a9be98e37d1940c944d0cca94c5ba185b67bc8a61ca", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 938.544, "unit": "us", "box": "gb10", "label": "screen_head_scan", "kernel": "screen_head::screen_head_scan", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 877.874, "pct_of_floor": 93.54, "bound": "memory", "status": "measured", "note": "verify LM head screen M=2 rows=51648 K=4096; n=246 calls, 2.0% of regime kernel time; peak class bf16; INT8 head copy + f32 scale/128, BF16 act, CUDA cores; median solo 938.54 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3106, 3119 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "242141bc0b2c3a679f5fba564abce2c519b1864def88fda409dc6dc093cf25ee", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 82.432, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "DSA q_absorb M=2 N=11264 K=1536", "floor_us": 69.858, "pct_of_floor": 84.75, "bound": "memory", "status": "measured", "note": "DSA q_absorb M=2 N=11264 K=1536; n=2709 calls, 1.9% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 81.70 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3121, 3134 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "c3a1be31317776cc34903ad7056dbe2febe4485068aed9a153557b10f6d78015", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 20.736, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16_batchm_few_fp32out", "kernel": "dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 9.55, "pct_of_floor": 46.06, "bound": "memory", "status": "measured", "note": "router M=2 N=288 K=4096; n=10343 calls, 1.9% of regime kernel time; peak class bf16; BF16 weight and act, f32 out; median solo 16.32 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3136, 3149 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "ab6edcc1be42224683f6930886ea8019f4b8e662ed1e41d7b6e914c09f5c5886", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 855.776, "unit": "us", "box": "gb10", "label": "dense_gemv_fp8w", "kernel": "dense_gemv_fp8w::dense_gemv_fp8w", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 850.399, "pct_of_floor": 99.37, "bound": "memory", "status": "measured", "note": "MTP draft head M=1 N=51619 K=4096; n=247 calls, 1.9% of regime kernel time; peak class bf16; W8A16 E4M3 per-row weight, BF16 act; median solo 855.78 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3151, 3164 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "022e3b54f2cedbd9296385505c8bbbecb479b6eb55319ae66af244a3ebeaf9a2", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 60.928, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16_batchm_few_fp32out_pair", "kernel": "dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out_pair", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 51.81, "pct_of_floor": 85.03, "bound": "memory", "status": "measured", "note": "DSA indexer wq_b M=2 N=4096 K=1536 + weights_proj N=32 K=4096; n=2709 calls, 1.4% of regime kernel time; peak class bf16; BF16 weight and act, f32 out; median solo 59.46 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3166, 3179 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "1579c117f9f693f8a09fb1655a203bf255fc79e302722cfc51c1cf24b39d0594", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 8.672, "unit": "us", "box": "gb10", "label": "glm5next_swiglu_clamp", "kernel": "glm5next_ffn::glm5next_swiglu_clamp", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "shared expert n=2x768", "floor_us": 0.037, "pct_of_floor": 0.43, "bound": "memory", "status": "measured", "note": "shared expert n=2x768; n=10343 calls, 1.2% of regime kernel time; peak class bf16; median solo 0.74 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3181, 3194 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "d310d98937536a388ca0ebebe52e107e49909d7854ad5313ae08c23c1a03c952", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 15.392, "unit": "us", "box": "gb10", "label": "kda_recurrent_decode_bf16_seq_reg_snap", "kernel": "kda_recurrent::kda_recurrent_decode_bf16_seq_reg_snap", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 17.688, "pct_of_floor": 114.92, "bound": "memory", "status": "measured", "note": "T=2 heads=22 dk=dv=128 1 snapshot; n=8375 calls, 1.1% of regime kernel time; peak class fp32; FP32 state read + written + 1 snapshot written; median solo 14.62 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3196, 3209 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "91ed5ebca7538159b0f687eb1cdcd12bd6085dc826d50d485cfbe5453c6b0479", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 15.36, "unit": "us", "box": "gb10", "label": "kda_o_norm_gated_bf16", "kernel": "kda_layer_ops::kda_o_norm_gated_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 0.182, "pct_of_floor": 1.18, "bound": "memory", "status": "measured", "note": "rows=2x22 head_dim=128; n=8375 calls, 1.1% of regime kernel time; peak class bf16; f32 in, BF16 gate/out; median solo 1.38 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3211, 3224 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "dda13cbce773c8bab39a84e3a9bc0bb38a9e320b43036473fcfc48e27adb7e3f", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 45.152, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8_split4", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8_split4", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 38.007, "pct_of_floor": 84.18, "bound": "memory", "status": "measured", "note": "DSA in-proj q_a|kv_a|idx wk|idx gate M=2 N=1536+512+128+128 K=4096; n=2709 calls, 1.1% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 43.74 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3226, 3239 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "59a9ef4142125740deea1e95e7ebd72f1cd702c9234a5c7fab30e0344a0db6b1", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 52.416, "unit": "us", "box": "gb10", "label": "w4a4_gemv_mx8", "kernel": "w4a4_gemv_mx::w4a4_gemv_mx8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 37.985, "pct_of_floor": 72.47, "bound": "memory", "status": "measured", "note": "dense FFN gate/up/down M=2 N=4096 K=4096; n=2223 calls, 1.0% of regime kernel time; peak class fp4; W4A4 NVFP4 x NVFP4 (static input scale); median solo 52.42 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3241, 3254 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "1e5eccd40ef31e497d49203ba13b80d7a8ab231d6c5f61351f0542ff9beb0725", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 7.872, "unit": "us", "box": "gb10", "label": "glm5next_router_topk_rank", "kernel": "glm5next_ffn::glm5next_router_topk_rank", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "rows=2 experts=288 top_k=8", "floor_us": 0.014, "pct_of_floor": 0.18, "bound": "memory", "status": "measured", "note": "rows=2 experts=288 top_k=8; n=10343 calls, 0.9% of regime kernel time; peak class bf16; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3256, 3269 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "a555a317557b7adb78958eed62a8b9e3a6fae6245be65e00985f986072259ef0", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 408.128, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "MTP o_absorb M=1 N=4096 K=11264", "floor_us": 370.704, "pct_of_floor": 90.83, "bound": "memory", "status": "measured", "note": "MTP o_absorb M=1 N=4096 K=11264; n=247 calls, 0.9% of regime kernel time; peak class bf16; BF16 weight and act; median solo 408.13 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3271, 3284 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "ca2e67025687765a9350c48bb42285d86ab074709311d0523874a1348282d9ab", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 9.376, "unit": "us", "box": "gb10", "label": "causal_conv1d_window_advance_to", "kernel": "kda_conv_tokens::causal_conv1d_window_advance_to", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 1.018, "pct_of_floor": 10.86, "bound": "memory", "status": "measured", "note": "snapshot T=1 dim=8448 d_conv=4; n=8375 calls, 0.7% of regime kernel time; peak class bf16; f32 conv window: 3 kept slots read, 4 written; median solo 3.39 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3286, 3299 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "beab3ec949fbd3150e3391672249829a25ceb5644ed8bc7bd0b17c9279486562", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 5.44, "unit": "us", "box": "gb10", "label": "w4a4_quant_rows_static", "kernel": "w4a4_gemv_mx_moe::w4a4_quant_rows_static", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 0.084, "pct_of_floor": 1.55, "bound": "memory", "status": "measured", "note": "rows=2 K=4096; n=11825 calls, 0.6% of regime kernel time; peak class bf16; BF16 in, E2M1 + E4M3/16 + f32 out; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3301, 3314 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "33222668aad5a02bbd5aa36a1565b81a81d4b715a3b8eab03589adb9b34b4085", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 7.776, "unit": "us", "box": "gb10", "label": "kda_gates_w8a8_k128_rows", "kernel": "kda_gates_w8a8::kda_gates_w8a8_k128_rows", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 3.217, "pct_of_floor": 41.38, "bound": "memory", "status": "measured", "note": "f_b|g_b M=2 N=2816+2816 K=128; n=8375 calls, 0.6% of regime kernel time; peak class fp8; W8A8 per-row E4M3 + gate/beta epilogues; median solo 7.78 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3316, 3329 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "58f52bf4ec483be242c92d607097a1725d99dac0a4d16e2783a87cc8cb668843", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 21.856, "unit": "us", "box": "gb10", "label": "glm5next_dsa_mla_decode_hb_fp8", "kernel": "glm5next_dsa_mla_decode_hb::glm5next_dsa_mla_decode_hb_fp8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "rows=2 heads=22 latent=512 splits=8", "floor_us": 3.812, "pct_of_floor": 17.44, "bound": "memory", "status": "measured", "note": "rows=2 heads=22 latent=512 splits=8; n=2709 calls, 0.5% of regime kernel time; peak class bf16; KV L=1646 (window midpoint: 1038 prompt + 51.1 tok/s x 11.9 s), 1646 selected tokens read once for both rows (rows' selections assumed to coincide), FP8 latent 512 B/token, q and o BF16; split-K partials (ws_o) not counted; median solo 21.86 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3331, 3344 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "5c74482a59a380c152945610a848ddb7ddd1c7a7429aa7f8cefe942364d3b34a", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 6.976, "unit": "us", "box": "gb10", "label": "kda_pack_qkv_bf16", "kernel": "kda_layer_ops::kda_pack_qkv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 0.271, "pct_of_floor": 3.89, "bound": "memory", "status": "measured", "note": "T=2 qkv=2816; n=8375 calls, 0.5% of regime kernel time; peak class bf16; copy; median solo 0.58 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3346, 3359 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "9662573fd1a98cba6b572ca7469368779ecf892ce1919f051a3f0920bdc817af", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 2.208, "unit": "us", "box": "gb10", "label": "glm5next_hc_post_rank_sum", "kernel": "glm5next_hc_post_sum::glm5next_hc_post_rank_sum", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 1.317, "pct_of_floor": 59.63, "bound": "memory", "status": "measured", "note": "T=2 ranks=3 hc=4 H=4096; n=22167 calls, 0.5% of regime kernel time; peak class bf16; 3 BF16 partials in, sum out, f32 residual streams in and out; median solo 2.21 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3361, 3374 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "6321267ea7ce06f16c577cd19b3c918c9b8a015688a5125a43938f06cc1e1ca8", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 294.704, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "MTP catch-up eh_proj M=1 N=4096 K=8192", "floor_us": 269.612, "pct_of_floor": 91.49, "bound": "memory", "status": "measured", "note": "MTP catch-up eh_proj M=1 N=4096 K=8192; n=172 calls, 0.4% of regime kernel time; peak class bf16; BF16 weight and act; median solo 294.70 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3376, 3389 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "d75c51909baff40e5e89c6c0fb349e18b1026f2e8731ace2afbd48aaf037a5c4", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 4.768, "unit": "us", "box": "gb10", "label": "causal_conv1d_update_l2norm_tokens", "kernel": "kda_conv_tokens::causal_conv1d_update_l2norm_tokens", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 1.086, "pct_of_floor": 22.77, "bound": "memory", "status": "measured", "note": "T=2 dim=8448 d_conv=4; n=8375 calls, 0.4% of regime kernel time; peak class bf16; state, input, BF16 taps, output; median solo 2.02 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3391, 3404 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "5a4ee11d7d993d958a4b674f467e9f5af110cb1603bab1f4479a64ca4049b4f7", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 3.456, "unit": "us", "box": "gb10", "label": "glm5next_moe_row_union", "kernel": "w4a16_gemv::glm5next_moe_row_union", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 0.001, "pct_of_floor": 0.03, "bound": "memory", "status": "measured", "note": "rows=2 x top_k=8 ids; n=10343 calls, 0.3% of regime kernel time; peak class bf16; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3406, 3419 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "d77d77c670775a2e0bfb90e6ee880559056e3fb36aa55cfe5aa864b57bf53e37", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 146.208, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "MTP q_absorb M=1 N=11264 K=1536", "floor_us": 139.071, "pct_of_floor": 95.12, "bound": "memory", "status": "measured", "note": "MTP q_absorb M=1 N=11264 K=1536; n=247 calls, 0.3% of regime kernel time; peak class bf16; BF16 weight and act; median solo 145.86 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3421, 3434 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "3b57068a194a7a362343e11d113e9d371d9b82e5357409d80c30773d9a34027a", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 65.776, "unit": "us", "box": "gb10", "label": "w4a16_gemv_sw_moe", "kernel": "w4a16_gemv::w4a16_gemv_sw_moe", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": "MTP routed gate or up slots=8 N=704 K=4096", "floor_us": 52.191, "pct_of_floor": 79.35, "bound": "memory", "status": "measured", "note": "MTP routed gate or up slots=8 N=704 K=4096; n=494 calls, 0.3% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act, 8 experts; median solo 61.30 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3436, 3449 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "3d1b5306c52eb123538d99ed72b6e028598315e9263f10322067e58421e0dfa9", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 190.688, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "shared expert down M=2 N=4096 K=768", "floor_us": 7.184, "pct_of_floor": 3.77, "bound": "memory", "status": "measured", "note": "shared expert down M=2 N=4096 K=768; n=9764 calls, 13.8% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; side stream, co-runs with the routed W4A4 sweep; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3450, 3463 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "605a0dbce007f2338b95158bb8c1604bbc258c82c9880d0bd2feb31402b4ca90", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 103.744, "unit": "us", "box": "gb10", "label": "glm5next_moe_combine", "kernel": "glm5next_ffn::glm5next_moe_combine", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "rows=2 top_k=8 H=4096", "floor_us": 0.658, "pct_of_floor": 0.63, "bound": "memory", "status": "measured", "note": "rows=2 top_k=8 H=4096; n=9764 calls, 9.0% of regime kernel time; peak class bf16; routed slots + shared in, BF16 out; median solo 1.44 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3465, 3478 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "f743326ffa786a0a7330da8b5acfc5c9454591c49ce9ce0f218b854aec5d7a67", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 90.208, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096", "floor_us": 80.948, "pct_of_floor": 89.73, "bound": "memory", "status": "measured", "note": "KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096; n=7906 calls, 6.4% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; median solo 85.15 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3480, 3493 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "071818dfd82a88cbf8fa3ca7a9cabc496bb4ee787b7f4a8a74453feb460d3245", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 199.648, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "DSA o_absorb M=2 N=4096 K=11264", "floor_us": 185.513, "pct_of_floor": 92.92, "bound": "memory", "status": "measured", "note": "DSA o_absorb M=2 N=4096 K=11264; n=2557 calls, 4.5% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 197.12 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3495, 3508 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "c248b17cadf49d8eae72ea0821ed49688834681716451e5bc30df2b7f46a3280", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 130.56, "unit": "us", "box": "gb10", "label": "dsa_index_scores_heads_window", "kernel": "dsa_scores_heads_window::dsa_index_scores_heads_window", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 67.561, "pct_of_floor": 51.75, "bound": "memory", "status": "measured", "note": "rows=2 heads=32 D=128 kpool=4; n=2557 calls, 2.9% of regime kernel time; peak class fp32; KV L=128653 (window midpoint: 128037 prompt + 51.8 tok/s x 11.9 s), P=32164 pools of f32 keys read once for both rows; median solo 130.56 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3510, 3523 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "8a13d8cff3f90f570baa2b42fba4bd754d8950145c00e6385ad4936201e9f896", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 34.432, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "KDA o_proj M=2 N=4096 K=2816", "floor_us": 26.168, "pct_of_floor": 76.0, "bound": "memory", "status": "measured", "note": "KDA o_proj M=2 N=4096 K=2816; n=7906 calls, 2.4% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; median solo 33.63 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3525, 3538 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "ad43ebab2b624b3125d5f0a1d88ad951070298e3bec8dc1974d3c22a817fe513", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 24.256, "unit": "us", "box": "gb10", "label": "w4a16_gemv_pipe_m2", "kernel": "w4a16_gemv_pipe::w4a16_gemv_pipe_m2", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "shared expert gate|up M=2 N=768+768 K=4096", "floor_us": 14.303, "pct_of_floor": 58.97, "bound": "memory", "status": "measured", "note": "shared expert gate|up M=2 N=768+768 K=4096; n=9764 calls, 2.1% of regime kernel time; peak class bf16; W4A16 NVFP4 weight (E2M1 + UE4M3/16), BF16 act; side stream; median solo 11.68 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3540, 3553 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "0b1bd2175d74dce33685bd995f646e578055669224eb0922280062a7e8040427", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 10.688, "unit": "us", "box": "gb10", "label": "glm5next_hc_pre_norm_c1", "kernel": "glm5next_mhc_c1::glm5next_hc_pre_norm_c1", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 3.851, "pct_of_floor": 36.03, "bound": "memory", "status": "measured", "note": "T=2 hc=4 H=4096 mix=24; n=20926 calls, 2.1% of regime kernel time; peak class bf16; BF16 hc_fn x f32 streams on CUDA cores; median solo 10.69 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3555, 3568 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "0b7bc1dd556935d3d1f2556352ce0bb26c5ad2b9f4826ea55482350384567f69", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 939.072, "unit": "us", "box": "gb10", "label": "screen_head_scan", "kernel": "screen_head::screen_head_scan", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 877.874, "pct_of_floor": 93.48, "bound": "memory", "status": "measured", "note": "verify LM head screen M=2 rows=51648 K=4096; n=232 calls, 1.9% of regime kernel time; peak class bf16; INT8 head copy + f32 scale/128, BF16 act, CUDA cores; median solo 939.07 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3570, 3583 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "8ea02ef3023ce555616239fddb56197c27bd47e143122aea332524e47cc3e8c1", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 81.664, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "DSA q_absorb M=2 N=11264 K=1536", "floor_us": 69.858, "pct_of_floor": 85.54, "bound": "memory", "status": "measured", "note": "DSA q_absorb M=2 N=11264 K=1536; n=2557 calls, 1.9% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 80.93 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3585, 3598 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "a78bae20d133ce328153e03decc38895b1d75d065f4aacdf7b4472bdd989ac86", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 871.392, "unit": "us", "box": "gb10", "label": "dense_gemv_fp8w", "kernel": "dense_gemv_fp8w::dense_gemv_fp8w", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 850.399, "pct_of_floor": 97.59, "bound": "memory", "status": "measured", "note": "MTP draft head M=1 N=51619 K=4096; n=233 calls, 1.8% of regime kernel time; peak class bf16; W8A16 E4M3 per-row weight, BF16 act; median solo 871.39 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3600, 3613 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "0534f44bf7d55f1b23f6f90f06020d027dd616f618e3395ccbd57a4bc68afecc", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 20.192, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16_batchm_few_fp32out", "kernel": "dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 9.55, "pct_of_floor": 47.3, "bound": "memory", "status": "measured", "note": "router M=2 N=288 K=4096; n=9764 calls, 1.8% of regime kernel time; peak class bf16; BF16 weight and act, f32 out; median solo 15.84 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3615, 3628 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "9524f9b27027708b76b8e6245c4b7c0a737b24620a42a70cf5ca322320f15c21", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 60.192, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16_batchm_few_fp32out_pair", "kernel": "dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out_pair", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 51.81, "pct_of_floor": 86.07, "bound": "memory", "status": "measured", "note": "DSA indexer wq_b M=2 N=4096 K=1536 + weights_proj N=32 K=4096; n=2557 calls, 1.4% of regime kernel time; peak class bf16; BF16 weight and act, f32 out; median solo 58.69 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3630, 3643 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "7f61f6cdf965624eb8f34d219a08c18766c75f37e8a1bcea4c310a89d86a7d7b", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 8.704, "unit": "us", "box": "gb10", "label": "glm5next_swiglu_clamp", "kernel": "glm5next_ffn::glm5next_swiglu_clamp", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "shared expert n=2x768", "floor_us": 0.037, "pct_of_floor": 0.43, "bound": "memory", "status": "measured", "note": "shared expert n=2x768; n=9764 calls, 1.1% of regime kernel time; peak class bf16; median solo 0.74 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3645, 3658 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "e6d49b5751bb6fa45e69ad23622caa70c5a8e36141e236b732782e91cebae7c8", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 15.84, "unit": "us", "box": "gb10", "label": "kda_recurrent_decode_bf16_seq_reg_snap", "kernel": "kda_recurrent::kda_recurrent_decode_bf16_seq_reg_snap", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 17.688, "pct_of_floor": 111.67, "bound": "memory", "status": "measured", "note": "T=2 heads=22 dk=dv=128 1 snapshot; n=7906 calls, 1.1% of regime kernel time; peak class fp32; FP32 state read + written + 1 snapshot written; median solo 15.04 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3660, 3673 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "5089f92ad9780cbac23078a240b558abd793cb2ef5701c35badc8363e881c57b", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 15.808, "unit": "us", "box": "gb10", "label": "kda_o_norm_gated_bf16", "kernel": "kda_layer_ops::kda_o_norm_gated_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 0.182, "pct_of_floor": 1.15, "bound": "memory", "status": "measured", "note": "rows=2x22 head_dim=128; n=7906 calls, 1.1% of regime kernel time; peak class bf16; f32 in, BF16 gate/out; median solo 1.38 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3675, 3688 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "5a6f4e539eee8b9609fc64683bed42f7d9915de2f270490c0f392cb39f75e4fb", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 44.96, "unit": "us", "box": "gb10", "label": "w8a8_gemv_rowscale_mb1_ku8_split4", "kernel": "w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8_split4", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 38.007, "pct_of_floor": 84.54, "bound": "memory", "status": "measured", "note": "DSA in-proj q_a|kv_a|idx wk|idx gate M=2 N=1536+512+128+128 K=4096; n=2557 calls, 1.0% of regime kernel time; peak class fp8; W8A8 per-row E4M3; median solo 43.58 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3690, 3703 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "c69efef2f0f394e15cd44c0cfafbfd5e4c40f07aa0d92e1d61a16347389e7178", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 52.384, "unit": "us", "box": "gb10", "label": "w4a4_gemv_mx8", "kernel": "w4a4_gemv_mx::w4a4_gemv_mx8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 37.985, "pct_of_floor": 72.51, "bound": "memory", "status": "measured", "note": "dense FFN gate/up/down M=2 N=4096 K=4096; n=2097 calls, 0.9% of regime kernel time; peak class fp4; W4A4 NVFP4 x NVFP4 (static input scale); median solo 52.38 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3705, 3718 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "82ba5b0c58c8882129cb22111abf75652a162745e38be196e033fa47e335d485", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 8.768, "unit": "us", "box": "gb10", "label": "glm5next_router_topk_rank", "kernel": "glm5next_ffn::glm5next_router_topk_rank", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "rows=2 experts=288 top_k=8", "floor_us": 0.014, "pct_of_floor": 0.16, "bound": "memory", "status": "measured", "note": "rows=2 experts=288 top_k=8; n=9764 calls, 0.9% of regime kernel time; peak class bf16; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3720, 3733 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "d7d174c6b43e36ba855664d778c9cd04d14e543b0da85a3fe8fc3f15ca39222e", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 407.392, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "MTP o_absorb M=1 N=4096 K=11264", "floor_us": 370.704, "pct_of_floor": 90.99, "bound": "memory", "status": "measured", "note": "MTP o_absorb M=1 N=4096 K=11264; n=233 calls, 0.8% of regime kernel time; peak class bf16; BF16 weight and act; median solo 407.39 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3735, 3748 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "1a484cd9381ea9623ca3035299095228a81fb97e25f71afec63895c3e1571989", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 32.48, "unit": "us", "box": "gb10", "label": "glm5next_dsa_mla_decode_hb_fp8", "kernel": "glm5next_dsa_mla_decode_hb::glm5next_dsa_mla_decode_hb_fp8", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "rows=2 heads=22 latent=512 splits=8", "floor_us": 4.645, "pct_of_floor": 14.3, "bound": "memory", "status": "measured", "note": "rows=2 heads=22 latent=512 splits=8; n=2557 calls, 0.7% of regime kernel time; peak class bf16; KV L=128653 (window midpoint: 128037 prompt + 51.8 tok/s x 11.9 s), 2051 selected tokens read once for both rows (rows' selections assumed to coincide), FP8 latent 512 B/token, q and o BF16; split-K partials (ws_o) not counted; median solo 32.48 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3750, 3763 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "c68817d06ed0c1e1540e476b22aaed94261c5bcc6c8fe4dc351de44c793ca439", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 9.152, "unit": "us", "box": "gb10", "label": "causal_conv1d_window_advance_to", "kernel": "kda_conv_tokens::causal_conv1d_window_advance_to", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 1.018, "pct_of_floor": 11.12, "bound": "memory", "status": "measured", "note": "snapshot T=1 dim=8448 d_conv=4; n=7906 calls, 0.6% of regime kernel time; peak class bf16; f32 conv window: 3 kept slots read, 4 written; median solo 3.26 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3765, 3778 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "47871f7265090b387ecb5059adf7557f228565f66455167dc252e0a16c9fdfb6", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 5.408, "unit": "us", "box": "gb10", "label": "w4a4_quant_rows_static", "kernel": "w4a4_gemv_mx_moe::w4a4_quant_rows_static", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 0.084, "pct_of_floor": 1.56, "bound": "memory", "status": "measured", "note": "rows=2 K=4096; n=11162 calls, 0.6% of regime kernel time; peak class bf16; BF16 in, E2M1 + E4M3/16 + f32 out; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3780, 3793 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "29cafbd3668456f303ba56e0913a598b8bd17b0251b03bb06d5b684f778b2db0", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 7.808, "unit": "us", "box": "gb10", "label": "kda_gates_w8a8_k128_rows", "kernel": "kda_gates_w8a8::kda_gates_w8a8_k128_rows", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 3.217, "pct_of_floor": 41.21, "bound": "memory", "status": "measured", "note": "f_b|g_b M=2 N=2816+2816 K=128; n=7906 calls, 0.6% of regime kernel time; peak class fp8; W8A8 per-row E4M3 + gate/beta epilogues; median solo 7.78 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3795, 3808 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "b34361bb0a866f829edd3b8f2b2032e4405e6e87be8f6db50e9ba4e5efbf1050", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 6.944, "unit": "us", "box": "gb10", "label": "kda_pack_qkv_bf16", "kernel": "kda_layer_ops::kda_pack_qkv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 0.271, "pct_of_floor": 3.91, "bound": "memory", "status": "measured", "note": "T=2 qkv=2816; n=7906 calls, 0.5% of regime kernel time; peak class bf16; copy; median solo 0.58 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3810, 3823 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "204b4a8231c0031957ea0834ac524ee2dde00b0c808add12d7b0a90cd17e3509", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 296.16, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "MTP catch-up eh_proj M=1 N=4096 K=8192", "floor_us": 269.612, "pct_of_floor": 91.04, "bound": "memory", "status": "measured", "note": "MTP catch-up eh_proj M=1 N=4096 K=8192; n=175 calls, 0.5% of regime kernel time; peak class bf16; BF16 weight and act; median solo 296.16 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3825, 3838 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "e8d7e9d566fdcac8e640e63588cf7c3a263915a59fa3369ea4901c40f5a4e442", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 2.464, "unit": "us", "box": "gb10", "label": "glm5next_hc_post_rank_sum", "kernel": "glm5next_hc_post_sum::glm5next_hc_post_rank_sum", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 1.317, "pct_of_floor": 53.43, "bound": "memory", "status": "measured", "note": "T=2 ranks=3 hc=4 H=4096; n=20925 calls, 0.5% of regime kernel time; peak class bf16; 3 BF16 partials in, sum out, f32 residual streams in and out; median solo 2.46 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3840, 3853 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "446835e6b59a3d64f3cfcbedba2bfd27010e8a09a819517197d4c0c80d3598cd", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 3.456, "unit": "us", "box": "gb10", "label": "glm5next_moe_row_union", "kernel": "w4a16_gemv::glm5next_moe_row_union", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 0.001, "pct_of_floor": 0.03, "bound": "memory", "status": "measured", "note": "rows=2 x top_k=8 ids; n=9764 calls, 0.3% of regime kernel time; peak class bf16; median solo 0.00 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3855, 3868 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "950c76bbcb268c4b87a641107fa39ce6a46d16778df0dd8489d511b4c66bca0d", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 4.672, "unit": "us", "box": "gb10", "label": "causal_conv1d_update_l2norm_tokens", "kernel": "kda_conv_tokens::causal_conv1d_update_l2norm_tokens", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": null, "floor_us": 1.086, "pct_of_floor": 23.24, "bound": "memory", "status": "measured", "note": "T=2 dim=8448 d_conv=4; n=7906 calls, 0.3% of regime kernel time; peak class bf16; state, input, BF16 taps, output; median solo 2.02 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3870, 3883 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "5d1c985e6cc70601f6586fcd206b01f3c3401658119b304372fff5c7b9ab9322", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 145.536, "unit": "us", "box": "gb10", "label": "dense_gemv_bf16", "kernel": "dense_gemv_bf16::dense_gemv_bf16", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u2248128k", "shape": "MTP q_absorb M=1 N=11264 K=1536", "floor_us": 139.071, "pct_of_floor": 95.56, "bound": "memory", "status": "measured", "note": "MTP q_absorb M=1 N=11264 K=1536; n=233 calls, 0.3% of regime kernel time; peak class bf16; BF16 weight and act; median solo 145.54 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3885, 3898 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p128k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "cbdd139ade2e0de9c0bc5897342b3d872053bb6b2f996061b53ab688e475a445", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 224.128, "unit": "us", "box": "gb10", "label": "w4a4_gemv_mx8_moe_union_sweep_sl_cs", "kernel": "w4a4_gemv_mx_moe::w4a4_gemv_mx8_moe_union_sweep_sl_cs", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 173.172, "pct_of_floor": 77.26, "bound": "memory", "status": "measured", "note": "routed gate+up union M=2x8 N=704+704 K=4096; n=10343 calls, 20.0% of regime kernel time; peak class fp4; D=13.854 union experts (measured: mean distinct experts of the 2-row verify union over 27972 layer-calls (666 verify launches x 42 MoE layers), METRALE_GLM_LAYER_DUMP moe_ids, same p1k prompt, binary dcaefeec (e586b7c3a), dgx1, 2026-10-22) x gate+up [704,4096] at 4.3125 bit/weight (scale dictionary; a tensor kept plain reads 4.5); median solo 214.94 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3900, 3913 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "11a901435a8e75e5730e2601c8b5fdb3fdefc81be62f40d2692dd8db9217cb10", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } }, { "benchmark": "kernel-perf", "metric": "kernel_time", "value": 104.0, "unit": "us", "box": "gb10", "label": "w4a4_gemv_mx8_moe_union_sweep_k64_cs", "kernel": "w4a4_gemv_mx_moe::w4a4_gemv_mx8_moe_union_sweep_k64_cs", "regime": "decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV\u22481k", "shape": null, "floor_us": 87.61, "pct_of_floor": 84.24, "bound": "memory", "status": "measured", "note": "routed down union slots=16 N=4096 K=704; n=10343 calls, 9.3% of regime kernel time; peak class fp4; D=13.854 union experts (measured: mean distinct experts of the 2-row verify union over 27972 layer-calls (666 verify launches x 42 MoE layers), METRALE_GLM_LAYER_DUMP moe_ids, same p1k prompt, binary dcaefeec (e586b7c3a), dgx1, 2026-10-22) x down [4096,704] at compact scales; slot activations stored 768 wide; median solo 102.56 us (after every earlier-launched kernel ended: PDL wait or concurrent fork excluded); TP=3 rank 0, W4A4 experts, w4a16 dense, w8a8 DSA, fp8_e4m3 KV, MTP 1 draft forced, shared fork 2; nsys --cuda-graph-trace=node, 8 s window of the warm repeat; kernel times include PDL/fork overlap waits (see solo time)", "source": { "file": "docs/kernel-perf/measurements.toml", "lines": [ 3915, 3928 ], "commit": "c8175076ded9f40709633697bd4eb26676d00467", "record": "nsys f13_p1k.sqlite @ c5b9a80e6 (fast13: binary f4735654, profile glm53-tp3-512k-fast13 = fast12 recipe), dgx1 rank 0, 2026-10-22" }, "fingerprint": { "record_sha256": "a32ff83ead39d9a27c36e7c33f7b41a6bd17c827bf2edc47fa84b7206b7c4ff9", "closure": null, "argv_sha256": null, "env_sha256": null, "draw_sha256": null, "git_sha": "c5b9a80e6" } } ], "evidence": [ { "kind": "circuit_plan", "id": "decode:1", "digest": "1ff131e588e38681729df0cb2ae65a4d9adc8cf91eeb184a1b8bfb7527ccb5fb" }, { "kind": "circuit_plan", "id": "multi_seq:2", "digest": "6901b4656acaaf7d12f09ccd77ab60bcaa1696dd7ebdcbb53b5f07bf99e2729f" }, { "kind": "circuit_plan", "id": "multi_seq:4", "digest": "97f9f6bb5b992cf3bd2a42dbdea856f941905452ad466044498902d8d6a10d9f" }, { "kind": "circuit_plan", "id": "multi_seq:8", "digest": "b7345097c70353d20e2ee062e090ed5993bc9bff45011cf32ad8726d7c13fdca" }, { "kind": "circuit_plan", "id": "multi_seq:12", "digest": "c30a839202ebafdcd5ce7478c5116ac40297816ce59385f79a0fd04cfd6d59cf" }, { "kind": "circuit_plan", "id": "multi_seq:16", "digest": "c39ed89d8b0fabba3184d6af97f6d1ed71fa74959bbfae24b3121ae5ac9cf13e" }, { "kind": "circuit_plan", "id": "multi_seq:24", "digest": "d242da1f6372bec8b7e0b36c0a46ed2d87d6fae1be2b65de557c1adf04a93e23" }, { "kind": "circuit_plan", "id": "multi_seq:32", "digest": "415a2b942fc2f0b326602c22e2fc36a1a385f949ea16490fb6d3f4a554134204" }, { "kind": "circuit_plan", "id": "multi_seq:48", "digest": "5ed8f0301f4e79e2a8dec20b6187fd37d9810cf563b4ca621466904d6264501b" }, { "kind": "circuit_plan", "id": "multi_seq:64", "digest": "ae9c63121d471006dab5fb16a6633c4c0a934d7534a8f9dc50abf519d7c6b228" }, { "kind": "circuit_plan", "id": "multi_seq:96", "digest": "e908fdfdccb6266d78c82148748a88673c46f83903903ba569026906c32c83fb" }, { "kind": "circuit_plan", "id": "multi_seq:128", "digest": "2a6b51a59e1f248d2972925af3219f37b5d5948d91b91c73371a9884c350626f" }, { "kind": "circuit_plan", "id": "verify:2", "digest": "5d2226cbcbc189e76766595702a1c403eadfb404408b147fa0215df818afa635" }, { "kind": "circuit_plan", "id": "verify:3", "digest": "3446f07c8c410b4e32cb9d065f8a6f2afa111cf9dedab48b6fd0d836be4fa4c6" }, { "kind": "circuit_plan", "id": "verify:4", "digest": "063961fa91fb7131b3bdd9ab41e9c8236a4307f474136c57f95a931fd836f38d" }, { "kind": "circuit_plan", "id": "verify:8", "digest": "3bc1c08e84d29eb84beb6e53b5dee5b74f195388ba6be44e394a0c66131a7f38" }, { "kind": "circuit_plan", "id": "draft:1", "digest": "8969b4c107e71640fdf1a199dcdc0ddc592ee3de57c33ba29a2fbe413a36ca62" } ], "provenance": { "kind": "preview", "engine_commit": "7f7fb57fc971b8d88f0fb08ab512321dd052f38c", "planner_sha256": "f4735654ee1d5a9bf4f56e836969140d5bfb26231104bf13118103c69ee8b156", "planner_built_from": "c5b9a80e635df7c1056a5b617272970f25daba64", "planner_header": { "recipe": "glm-5.3/glm-5.3-flash-nvfp4-tp3-512k", "checkpoint": "nvidia/GLM-5.3-Flash-NVFP4", "target": "gb10/glm-5.3-flash/nvfp4", "settings": "activation_quantization=adaptive decode_split_silu=on gemv_sw=on glm_mhc_fused=on glm_moe_expert_layout=tp glm_moe_w4a4_grouped=on glm_router_rows=on glm_router_topk_rank=on kv_cache_dtype=fp8 lm_head_dtype=bf16 rms_norm_act_quant=off row_tiers=by_rows ssm_ba_gates_hopper=on ssm_batched_recurrent=on ssm_h_dtype=f32 w4a16_tc=on", "opt-in levers": "none" }, "commands": [ "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode decode --rows 1", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 2", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 4", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 8", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 12", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 16", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 24", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 32", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 48", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 64", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 96", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode multi_seq --rows 128", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode verify --rows 2", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode verify --rows 3", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode verify --rows 4", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode verify --rows 8", "met circuit plan --root --checkpoint glm-5.3/glm-5.3-flash-nvfp4-tp3-512k --hardware gb10 --precision recipe --format plan --mode draft --rows 1" ], "files": [ { "path": "kernels/circuits/INSTANCES.toml", "sha256": "126b177aa899733c7f2909a448bfbfc45c3a97588c31205c405aa8bea6fadb7d" }, { "path": "recipes/glm-5.3/glm-5.3-flash-nvfp4-tp3-512k.yaml", "sha256": "6e639ddb73e534fa93c43543837dfd30472eea3f377f972ef3de99a97bb22827" }, { "path": "kernels/gb10/HARDWARE.toml", "sha256": "8a787d560510594f9694a85836879cc9df1c55d874f8fcf37f0701aac4e63a12" }, { "path": "kernels/DEVICES.toml", "sha256": "256d85840f0e77ec84c5fc739f30aacff6bf24775cc8bef344e36647e51e9c25" }, { "path": "kernels/circuits/glm5_next.toml", "sha256": "047893d2422df5acbcaac36039e4f0655b679df7ead3265d022139f013b14a81" }, { "path": "kernels/circuits/blocks/glm5_next_kda.toml", "sha256": "9486bdea4533af4d30c951c449e39037d639bd791ff66a70f62d3d4e6d44b795" }, { "path": "kernels/circuits/blocks/glm5_next_dsa.toml", "sha256": "fdb3145558a6c1b7f9256c34854ae39114868d534e4f0062d91688fa91fa5eef" }, { "path": "kernels/circuits/blocks/glm5_next_ffn.toml", "sha256": "4738cd396a115447da27fe1ac4f0d1d324f4ac595af6a98b095231d62b698110" }, { "path": "kernels/gb10/common/KERNEL_FAMILIES.toml", "sha256": "70c24b8b47644bdbe040157cc1604f99a8a143401f44b4282620737123ac8571" }, { "path": "kernels/circuits/precision/glm-5.3-flash-nvfp4-tp3.toml", "sha256": "f1e299043e34000299680b8d2fce60b62342ccc487d9cb9734c9ddc7dca6d15a" } ] }, "settings": { "activation_quantization": "adaptive", "decode_split_silu": "on", "gemv_sw": "on", "glm_mhc_fused": "on", "glm_moe_expert_layout": "tp", "glm_moe_w4a4_grouped": "on", "glm_router_rows": "on", "glm_router_topk_rank": "on", "kv_cache_dtype": "fp8", "lm_head_dtype": "bf16", "rms_norm_act_quant": "off", "row_tiers": "by_rows", "ssm_ba_gates_hopper": "on", "ssm_batched_recurrent": "on", "ssm_h_dtype": "f32", "w4a16_tc": "on" }, "target": "gb10/glm-5.3-flash/nvfp4" }