chore(data): add matched B200 SGLang 0.5.14 power by kaim-eng · Pull Request #1535 · ai-dynamo/aiconfigurator

kaim-eng · 2026-08-13T15:20:49Z

Important

Dependency and zero-sentinel update

This pure-data PR depends on #1590. Merge #1590 first, then rebase/update this PR onto the resulting main before merging it.

Latest signed-off data head: 6bc4f243cb5c7bf20a4e8e27d90819a41f186d93.

The latest commit normalizes every unavailable power / power_limit cell in the changed power-enabled tables to typed 0.0. Row coverage, measured values, logical keys, and physical row order are unchanged.

Summary

Pure-data PR constructed from the then-current main data snapshot (095f58a51c4ca8e61b66ec108d86f223f8d559ce). This revision rebuilds the B200 SGLang 0.5.14 power overlay from the third-pass campaign pinned to the same backend image used for the on-stock perf-only collection.

updates exactly 24 B200 SGLang 0.5.14 parquet files in the aggregate PR diff
retains the complete stock row set and physical row order
preserves the schema/type/value of every original field
appends non-null float64 power / power_limit only through exact logical-key matches
retains unmatched stock rows as typed 0.0 / 0.0 sentinel pairs and excludes collection-only keys

No code, documentation, sidecar, or other artifact is included in this PR.

Exact-image provenance

on-stock perf-only snapshot: cb43a2f412983fee709c087882ea3438ab5cb8f5
collector code used for power collection: fix(collector): preserve power data across B200 collection paths #1534 head 0c6d722ecbbbce39c477adceae2687eb7245d22c
image: lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
local image (.sqsh) SHA-256: b3ecf5c8fcd521dcca67ceaa4c786d0b73b6adabdc5e08885c092b7c64c4551c
SGLang source revision in the image: 49e384ce9d304648e9959666ecb8ce8cd98d0deb

Broad collection passes used umb-b200-* 8x-B200 allocations and parallelized independent single-GPU cases across the available GPUs. Final exact-key retries used isolated 1-GPU allocations where the operation did not require multiple GPUs; the custom-allreduce collection retained its required multi-GPU topology.

The only approved key normalization is the one-to-one HCA model alias sgl-project/DeepSeek-V4-Pro-FP8 -> deepseek-ai/DeepSeek-V4-Pro for the two HCA tables. No performance field is relabeled or changed.

Coverage for the packaged stock tables

stock rows retained across the 24 changed tables: 624,463 / 624,463
rows with complete measured power pairs: 575,908
rows remaining perf-only: 48,555
raw row coverage: 92.2245%
partial power / power_limit pairs: 0

Twenty-one files are byte-identical to the prior PR head. Three tables changed on main after the original collection snapshot, so their power was rejoined by the complete logical key while preserving every current-main row, field, latency value, type, and physical row order:

Table	Stock rows	Exact-key powered rows	Unmeasured `0.0` pairs
`moe_perf.parquet`	151,470	145,314	6,156
`dsa_context_module_perf.parquet`	95,203	55,255	39,948
`dsa_generation_module_perf.parquet`	6,048	3,600	2,448

No power was inferred for new or changed current-main keys that were absent from the collection snapshot.

Overlay policy

The output is constructed from the stock tables, then power is joined by the complete logical key. Later isolated retries take precedence only for the same exact key. This makes the stock support surface authoritative: collected-only configurations are not inserted, and stock latency or configuration values cannot be overwritten.

Power-quality checks

every measured power pair is finite
every measured pair has power > 0 and power_limit > 0; unavailable pairs are exactly 0.0 / 0.0
905 measurements exceed the recorded 1000 W limit briefly; the maximum is 1.027278x the limit (+2.7278%), within the approved 5% transient allowance
duplicate stock logical keys have consistent power pairs

Rebase and validation

overlay construction base: then-current main data snapshot 095f58a51c4ca8e61b66ec108d86f223f8d559ce; rebase onto the post-fix(collector): normalize missing power metrics to zero #1590 main is required before merge
current signed-off head: 6bc4f243cb5c7bf20a4e8e27d90819a41f186d93
history: two signed-off data-only commits
diff: exactly 24 B200 SGLang 0.5.14 parquet files; no code, documentation, sidecar, or legacy perf artifact
independent Arrow audit: all 624,463 stock rows, original columns, values, types, latency, and physical order preserved
exact-key transfer: 575,908 complete measured power pairs, 48,555 typed 0.0 sentinel pairs, 0 null or partial pairs
strict repository parquet diff: PASS; +0 / -0 / ~575,908 rows, 0 legacy perf files

The PR is structurally independent of #1534. Targeted support-matrix validation with #1590 produced 118 PASS, 2 HYBRID_PASS, and 22 unrelated FAIL rows, with zero null-access errors.

copy-pr-bot · 2026-08-13T15:20:56Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

coderabbitai · 2026-08-13T15:21:03Z

Important

Review skipped

Review was skipped due to path filters

⛔ Files ignored due to path filters (24)

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/context_attention_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/generation_attention_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/comm/sglang/0.5.14/custom_allreduce_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/encoder_attention/sglang/0.5.14/encoder_attention_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/gemm/sglang/0.5.14/gemm_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/linear_attention/sglang/0.5.14/gdn_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mhc/sglang/0.5.14/mhc_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/context_mla_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/generation_mla_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla_bmm/sglang/0.5.14/mla_bmm_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/moe/sglang/0.5.14/moe_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/computescale_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/scale_matrix_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_context_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_generation_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_context_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_generation_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_topk_calib_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_context_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_generation_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_paged_mqa_logits_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_dsa_attn_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_mqa_logits_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**
aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_topk_module_perf.parquet is excluded by !**/*.parquet and included by aic-core/**

CodeRabbit blocks several paths by default. You can override this behavior by explicitly including those paths in the path filters. For example, including **/dist/** will override the default block on the dist directory, by removing the pattern from both the lists.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: eb02eea9-9410-42a1-b625-bd16ec086a7a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

_{Comment @coderabbitai help to get the list of available commands.}

github-actions · 2026-08-13T15:22:36Z

Perf Parquet Diff Report

Artifact bundle: perf-parquet-diff (retained for 30 days)

Compared origin/main to HEAD for aic-core/src/aiconfigurator_core/systems/data.

Parquet files changed: 24
CSV-to-parquet conversions checked: 0
Conversions with matching columns and rows: 0
New parquet files without a base CSV/parquet counterpart: 0
Modified or renamed parquet files: 24
Deleted parquet files: 0
Legacy *_perf.txt files added or modified: 0
Row-level changes: +0 / -0 / ~624463
Invalid power-metric files: 0
Full per-file diff artifacts: 24 files under parquet-diff-details/diffs/

Per-File Row Diff Preview

Showing the first 3 rows per diff kind for each changed parquet file. Full exact CSVs are in parquet-diff-details/.

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/context_attention_perf.parquet

Rows: +0 / -0 / ~43211
Key columns: framework, version, device, op_name, kernel_source, batch_size, isl, num_heads, num_key_value_heads, head_dim, beam_width, attn_dtype, kv_cache_dtype, step, window_size

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/context_attention_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,batch_size,isl,num_heads,num_key_value_heads,head_dim,beam_width,attn_dtype,kv_cache_dtype,step,window_size,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,context_attention,flashinfer,1,1,1,1,128,1,bfloat16,bfloat16,0,0,,305.366,,1000.0
SGLang,0.5.14,NVIDIA B200,context_attention,flashinfer,1,1,1,1,128,1,bfloat16,fp8,0,0,,300.529,,1000.0
SGLang,0.5.14,NVIDIA B200,context_attention,flashinfer,1,1,16,1,128,1,bfloat16,bfloat16,0,0,,385.902,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/generation_attention_perf.parquet

Rows: +0 / -0 / ~48501
Key columns: framework, version, device, op_name, kernel_source, batch_size, isl, num_heads, num_key_value_heads, head_dim, beam_width, attn_dtype, kv_cache_dtype, step, window_size

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/sglang/0.5.14/generation_attention_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,batch_size,isl,num_heads,num_key_value_heads,head_dim,beam_width,attn_dtype,kv_cache_dtype,step,window_size,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,generation_attention,flashinfer,1,1,1,1,128,1,bfloat16,bfloat16,1,0,,668.08,,1000.0
SGLang,0.5.14,NVIDIA B200,generation_attention,flashinfer,1,1,1,1,128,1,bfloat16,bfloat16,1023,0,,707.142,,1000.0
SGLang,0.5.14,NVIDIA B200,generation_attention,flashinfer,1,1,1,1,128,1,bfloat16,bfloat16,127,0,,543.522,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/comm/sglang/0.5.14/custom_allreduce_perf.parquet

Rows: +0 / -0 / ~138
Key columns: framework, version, device, op_name, kernel_source, allreduce_dtype, num_gpus, message_size, backend

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/comm/sglang/0.5.14/custom_allreduce_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,allreduce_dtype,num_gpus,message_size,backend,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,all_reduce,SGLang_CustomAllReduce_eager,bfloat16,2,1024,sglang_eager,,255.336,,1000.0
SGLang,0.5.14,NVIDIA B200,all_reduce,SGLang_CustomAllReduce_eager,bfloat16,2,1048576,sglang_eager,,331.051,,1000.0
SGLang,0.5.14,NVIDIA B200,all_reduce,SGLang_CustomAllReduce_eager,bfloat16,2,128,sglang_eager,,252.488,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/encoder_attention/sglang/0.5.14/encoder_attention_perf.parquet

Rows: +0 / -0 / ~7679
Key columns: framework, version, device, op_name, kernel_source, batch_size, isl, num_heads, head_dim, attn_dtype

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/encoder_attention/sglang/0.5.14/encoder_attention_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,batch_size,isl,num_heads,head_dim,attn_dtype,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,encoder_attention,flash_attention_v4,1,1,1,64,bfloat16,,958.84,,1000.0
SGLang,0.5.14,NVIDIA B200,encoder_attention,flash_attention_v4,1,1,1,72,bfloat16,,992.717,,1000.0
SGLang,0.5.14,NVIDIA B200,encoder_attention,flash_attention_v4,1,1,10,64,bfloat16,,985.955,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/gemm/sglang/0.5.14/gemm_perf.parquet

Rows: +0 / -0 / ~130536
Key columns: framework, version, device, op_name, kernel_source, gemm_dtype, m, n, k

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/gemm/sglang/0.5.14/gemm_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,gemm_dtype,m,n,k,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,gemm,sglang_deepgemm_gemm_nt_f8f8bf16,fp8_block,1,1024,1024,,641.5605,,1000.0
SGLang,0.5.14,NVIDIA B200,gemm,sglang_deepgemm_gemm_nt_f8f8bf16,fp8_block,1,1024,10240,,581.8160000000001,,1000.0
SGLang,0.5.14,NVIDIA B200,gemm,sglang_deepgemm_gemm_nt_f8f8bf16,fp8_block,1,1024,12288,,830.154,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/linear_attention/sglang/0.5.14/gdn_perf.parquet

Rows: +0 / -0 / ~8820
Key columns: framework, version, device, op_name, kernel_source, phase, batch_size, seq_len, num_tokens, d_model, d_conv, num_k_heads, head_k_dim, num_v_heads, head_v_dim, model_name

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/linear_attention/sglang/0.5.14/gdn_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,phase,batch_size,seq_len,num_tokens,d_model,d_conv,num_k_heads,head_k_dim,num_v_heads,head_v_dim,model_name,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,16,128,16,128,Qwen/Qwen3.5-0.8B,,246.104,,1000.0
SGLang,0.5.14,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,2,128,2,128,Qwen/Qwen3.5-0.8B,,250.571,,1000.0
SGLang,0.5.14,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,4,128,4,128,Qwen/Qwen3.5-0.8B,,241.958,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mhc/sglang/0.5.14/mhc_module_perf.parquet

Rows: +0 / -0 / ~140
Key columns: framework, version, device, op_name, kernel_source, architecture, num_tokens, num_sites, hc_mult, hidden_size, sinkhorn_iters

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mhc/sglang/0.5.14/mhc_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,architecture,num_tokens,num_sites,hc_mult,hidden_size,sinkhorn_iters,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,post,sglang_tilelang_mhc_post,DeepseekV4ForCausalLM,1,2,4,4096,20,,244.082,,1000.0
SGLang,0.5.14,NVIDIA B200,post,sglang_tilelang_mhc_post,DeepseekV4ForCausalLM,1,2,4,7168,20,,243.281,,1000.0
SGLang,0.5.14,NVIDIA B200,post,sglang_tilelang_mhc_post,DeepseekV4ForCausalLM,1024,2,4,4096,20,,304.426,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/context_mla_perf.parquet

Rows: +0 / -0 / ~1760
Key columns: framework, version, device, op_name, kernel_source, mla_dtype, kv_cache_dtype, num_heads, batch_size, isl, tp_size, step

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/context_mla_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,mla_dtype,kv_cache_dtype,num_heads,batch_size,isl,tp_size,step,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,mla_context,trtllm_mla,bfloat16,bfloat16,1,1,1,64,0,,256.178,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_context,trtllm_mla,bfloat16,bfloat16,1,1,1024,64,0,,405.908,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_context,trtllm_mla,bfloat16,bfloat16,1,1,10240,64,0,,700.2407999999999,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/generation_mla_perf.parquet

Rows: +0 / -0 / ~2896
Key columns: framework, version, device, op_name, kernel_source, mla_dtype, kv_cache_dtype, num_heads, batch_size, isl, tp_size, step

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/sglang/0.5.14/generation_mla_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,mla_dtype,kv_cache_dtype,num_heads,batch_size,isl,tp_size,step,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,mla_generation,trtllm_mla,bfloat16,bfloat16,1,1,1,64,1,,249.116,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_generation,trtllm_mla,bfloat16,bfloat16,1,1,1,64,1023,,522.559,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_generation,trtllm_mla,bfloat16,bfloat16,1,1,1,64,127,,346.071,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla_bmm/sglang/0.5.14/mla_bmm_perf.parquet

Rows: +0 / -0 / ~1272
Key columns: framework, version, device, op_name, kernel_source, bmm_dtype, num_tokens, num_heads

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla_bmm/sglang/0.5.14/mla_bmm_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,bmm_dtype,num_tokens,num_heads,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,mla_gen_post,sglang_sgl_kernel_bmm_fp8,fp8,1,1,,243.905,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_gen_post,sglang_sgl_kernel_bmm_fp8,fp8,1,12,,243.714,,1000.0
SGLang,0.5.14,NVIDIA B200,mla_gen_post,sglang_sgl_kernel_bmm_fp8,fp8,1,128,,243.329,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/moe/sglang/0.5.14/moe_perf.parquet

Rows: +0 / -0 / ~151470
Key columns: framework, version, device, op_name, kernel_source, moe_dtype, num_tokens, hidden_size, inter_size, topk, num_experts, moe_tp_size, moe_ep_size, distribution

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/moe/sglang/0.5.14/moe_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,moe_dtype,num_tokens,hidden_size,inter_size,topk,num_experts,moe_tp_size,moe_ep_size,distribution,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,moe,sglang_flashinfer_trtllm_moe,bfloat16,1,2048,512,8,256,1,1,balanced,,688.924,,1000.0
SGLang,0.5.14,NVIDIA B200,moe,sglang_flashinfer_trtllm_moe,bfloat16,1,2048,512,8,256,1,1,power_law_1.01,,490.709,,1000.0
SGLang,0.5.14,NVIDIA B200,moe,sglang_flashinfer_trtllm_moe,bfloat16,1,2048,512,8,256,1,1,power_law_1.2,,458.592,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/computescale_perf.parquet

Rows: +0 / -0 / ~1628
Key columns: framework, version, device, op_name, kernel_source, m, k, quant_dtype

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/computescale_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,m,k,quant_dtype,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,compute_scale,sglang,1,1024,fp8,,264.515,,1000.0
SGLang,0.5.14,NVIDIA B200,compute_scale,sglang,1,10240,fp8,,256.113,,1000.0
SGLang,0.5.14,NVIDIA B200,compute_scale,sglang,1,12288,fp8,,264.819,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/scale_matrix_perf.parquet

Rows: +0 / -0 / ~1628
Key columns: framework, version, device, op_name, kernel_source, m, k, quant_dtype

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/sglang/0.5.14/scale_matrix_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,m,k,quant_dtype,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,scale_matrix,sglang,1,1024,fp8,,264.738,,1000.0
SGLang,0.5.14,NVIDIA B200,scale_matrix,sglang,1,10240,fp8,,256.129,,1000.0
SGLang,0.5.14,NVIDIA B200,scale_matrix,sglang,1,12288,fp8,,264.294,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_context_module_perf.parquet

Rows: +0 / -0 / ~95203
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_context_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsa_context_module,sglang_dsa_dense_mha_trtllm_ragged,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1024,1,0,,585.1595,,1000.0
SGLang,0.5.14,NVIDIA B200,dsa_context_module,sglang_dsa_dense_mha_trtllm_ragged,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1024,1,1,,551.1144,,1000.0
SGLang,0.5.14,NVIDIA B200,dsa_context_module,sglang_dsa_dense_mha_trtllm_ragged,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1024,1,1024,,494.4267,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_generation_module_perf.parquet

Rows: +0 / -0 / ~6048
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsa_generation_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsa_generation_module,sglang_dsa_indexer_trtllm,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1,1,1,,327.728,,1000.0
SGLang,0.5.14,NVIDIA B200,dsa_generation_module,sglang_dsa_indexer_trtllm,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1,1,10000,,356.21,,1000.0
SGLang,0.5.14,NVIDIA B200,dsa_generation_module,sglang_dsa_indexer_trtllm,deepseek-ai/DeepSeek-V3.2,DeepseekV32ForCausalLM,bfloat16,bfloat16,fp8_block,128,1,1,1,1024,,304.919,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_context_module_perf.parquet

Rows: +0 / -0 / ~43804
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_context_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsv4_csa_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,0,4,,326.275,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_csa_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1,4,,479.72166666666664,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_csa_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,10000,4,,463.895,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_generation_module_perf.parquet

Rows: +0 / -0 / ~3200
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_generation_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsv4_csa_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1,4,,244.348,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_csa_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1024,4,,534.0245,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_csa_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,10240,4,,471.00600000000003,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_topk_calib_perf.parquet

Rows: +0 / -0 / ~10364
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio, score_mode

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_csa_topk_calib_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,score_mode,power__base,power__head,power_limit__base,power_limit__head
SGLang,kernel-level,NVIDIA B200,dsv4_csa_topk_calib,topk_transform_v1,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,128,1,1,1,0,4,v1_flat,,0.0,,1000.0
SGLang,kernel-level,NVIDIA B200,dsv4_csa_topk_calib,topk_transform_v1,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,128,1,1,1,0,4,v1_top_last,,0.0,,1000.0
SGLang,kernel-level,NVIDIA B200,dsv4_csa_topk_calib,topk_transform_v1,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,128,1,1,1,1,4,v1_flat,,0.0,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_context_module_perf.parquet

Rows: +0 / -0 / ~47456
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_context_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsv4_hca_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,0,128,,372.981,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_hca_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1,128,,543.88725,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_hca_context_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,10000,128,,510.95425,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_generation_module_perf.parquet

Rows: +0 / -0 / ~3200
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_hca_generation_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,0.5.14,NVIDIA B200,dsv4_hca_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1,128,,245.1135,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_hca_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,1024,128,,279.623,,1000.0
SGLang,0.5.14,NVIDIA B200,dsv4_hca_generation_module,compressed_flashmla,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8_e4m3,bfloat16,128,1,1,1,10240,128,,297.8966666666667,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_paged_mqa_logits_module_perf.parquet

Rows: +0 / -0 / ~2181
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/dsv4_paged_mqa_logits_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,kernel-level,NVIDIA B200,dsv4_paged_mqa_logits_module,deep_gemm.fp8_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,0,4,,243.179,,1000.0
SGLang,kernel-level,NVIDIA B200,dsv4_paged_mqa_logits_module,deep_gemm.fp8_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1,4,,244.755,,1000.0
SGLang,kernel-level,NVIDIA B200,dsv4_paged_mqa_logits_module,deep_gemm.fp8_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,10000,4,,250.794,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_dsa_attn_module_perf.parquet

Rows: +0 / -0 / ~3332
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_dsa_attn_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,kernel-level,NVIDIA B200,glm5_dsa_attn_module,flash_mla_sparse_fwd,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,bfloat16,fp8_e4m3,fp8_block,64,1,1,1,1,1,,942.4306666666666,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_dsa_attn_module,flash_mla_sparse_fwd,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,bfloat16,fp8_e4m3,fp8_block,64,1,1,1,1024,1,,272.5563333333333,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_dsa_attn_module,flash_mla_sparse_fwd,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,bfloat16,fp8_e4m3,fp8_block,64,1,1,1,1048575,1,,277.932,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_mqa_logits_module_perf.parquet

Rows: +0 / -0 / ~3332
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_mqa_logits_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
SGLang,kernel-level,NVIDIA B200,glm5_mqa_logits_module,deep_gemm.fp8_mqa_logits,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1,1,,922.168,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_mqa_logits_module,deep_gemm.fp8_mqa_logits,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1024,1,,571.239,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_mqa_logits_module,deep_gemm.fp8_mqa_logits,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1048575,1,,280.77840000000003,,1000.0

aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_topk_module_perf.parquet

Rows: +0 / -0 / ~6664
Key columns: framework, version, device, op_name, kernel_source, model, architecture, mla_dtype, kv_cache_dtype, gemm_type, num_heads, batch_size, isl, tp_size, step, compress_ratio, score_mode

modified rows - full CSV: parquet-diff-details/aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/sglang/0.5.14/glm5_topk_module_perf.parquet.modified.csv

framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,batch_size,isl,tp_size,step,compress_ratio,score_mode,power__base,power__head,power_limit__base,power_limit__head
SGLang,kernel-level,NVIDIA B200,glm5_topk_module,fast_topk_transform_fused,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1,1,flat,,0.0,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_topk_module,fast_topk_transform_fused,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1,1,top_last,,0.0,,1000.0
SGLang,kernel-level,NVIDIA B200,glm5_topk_module,fast_topk_transform_fused,nvidia/GLM-5.2-NVFP4,GlmMoeDsaForCausalLM,fp8_e4m3,fp8_e4m3,fp8_block,64,1,1,1,1024,1,flat,,0.0,,1000.0

Artifact Contents

Full per-file unified diffs: perf-parquet-diff artifact under parquet-diff-details/diffs/ (24 files)
Exact row-level CSVs: perf-parquet-diff artifact under parquet-diff-details/ (listed in summary.csv)

github-actions · 2026-08-13T15:24:08Z

Sanity Check Chart Generation Report

📥 Download all sanity charts from workflow artifacts

New perf data files were detected in this PR. Please use the link above to
download sanity check charts for the new perf data to compare the collected
perf data vs SOL (theoretical max performance).

Below is a report of whether the chart generation was successful for each op.
If doesn't validate whether the perf data itself is sane.

Chart Generation Report for system: b200_sxm, backend: sglang, backend_version: 0.5.14

gemm ✅
context_attention ✅
context_attention_with_prefix ✅
generation_attention ✅
generation_attention_b ✅
context_mla_with_prefix ✅
generation_mla ✅
generation_mla_b ✅
moe ✅
allreduce ✅
dsa_module ✅
CLI smoke test ✅

Signed-off-by: Kai Ma <kaim@nvidia.com>

Signed-off-by: kaim-eng <kaim@NV-8QHBYK4.localdomain>

tianhaox · 2026-09-02T07:12:47Z

Independent Arrow audit of the 24 files at c8be175 against the merge-base (2ed278a, the #1590 merge): every non-power column is value- and order-identical to stock, row counts unchanged, power/power_limit are typed float64 with 0 nulls, max power/limit ratio 1.0273, no duplicate-key power inconsistencies. Two things need fixing before this can merge.

1. 7,966 partial pairs (power=0.0, power_limit=1000.0) in the two topk tables.

table	rows	`power>0 & limit>0`	`power==0 & limit==1000`	`0.0/0.0`
`dsv4_csa_topk_calib_perf`	10,364	4,842	5,522	0
`glm5_topk_module_perf`	6,664	4,220	2,444	0

They are exactly the stock rows with latency == 0.0 (5,522 + 2,444). The PR contract says unavailable pairs are 0.0 / 0.0, and the body reports "partial pairs: 0" and 575,908 measured rows; the actual measured count is 567,942 (90.95%), the 7,966 partials were counted as measured. Runtime-safe (both loaders read only power, so these rows already resolve as uncovered), but please normalize them to 0.0 / 0.0 and refresh the numbers. Side note: #1590's parquet_diff gate only checks type/null/finite/non-negative, so it cannot catch this shape; a pair-consistency check would be a cheap follow-up.

2. The rebase onto current main is not mechanical for gemm_perf.

#1533 (b31d899) re-collected the SGLang 0.5.14 fp8_block DeepGEMM rows after moving the UE8M0 weight-scale pack out of the timed loop. Against this PR's stock: 29,526 rows have new latency on main, 26,071 faster, median |drift| 32.9%, 18,442 rows above 25%. The power here was collected at 0c6d722 with the old timed loop, so it describes a different kernel mix than the latency it would be joined to. Please apply the same 25% per-row drift gate used in #1536/#1584 (those rows fall to 0.0 / 0.0), or re-collect GEMM power with the post-#1533 collector.

moe_perf is the easy one: main added 3,159 nvfp4 rows (#1519), zero latency changes on the 151,470 surviving keys, so a plain key re-overlay gives 145,314 / 154,629 powered and the new rows get sentinels.

kaim-eng changed the title ~~data: add matched B200 SGLang 0.5.14 power~~ chore(data): add matched B200 SGLang 0.5.14 power Aug 13, 2026

github-actions Bot added the chore label Aug 13, 2026

kaim-eng force-pushed the codex/b200-sglang-0.5.14-power-data branch 2 times, most recently from 1f47c48 to 130f081 Compare August 13, 2026 16:59

kaim-eng mentioned this pull request Aug 14, 2026

fix(collector): preserve power data across B200 collection paths #1534

Open

kaim-eng marked this pull request as ready for review August 16, 2026 13:04

kaim-eng requested review from a team as code owners August 16, 2026 13:04

kaim-eng force-pushed the codex/b200-sglang-0.5.14-power-data branch 2 times, most recently from 4b85114 to 3dee700 Compare August 21, 2026 13:52

kaim-eng requested review from a team as code owners August 21, 2026 13:52

kaim-eng changed the base branch from codex/b200-power-collector-fixes to main August 21, 2026 13:52

kaim-eng force-pushed the codex/b200-sglang-0.5.14-power-data branch from 1088503 to 6bc4f24 Compare August 24, 2026 12:59

kaim-eng mentioned this pull request Aug 24, 2026

fix(collector): normalize missing power metrics to zero #1590

Merged

kaim-eng and others added 2 commits August 26, 2026 10:51

chore(data): add matched B200 SGLang 0.5.14 power

0205919

Signed-off-by: Kai Ma <kaim@nvidia.com>

fix(data): normalize missing SGLang power metrics

c8be175

Signed-off-by: kaim-eng <kaim@NV-8QHBYK4.localdomain>

kaim-eng force-pushed the codex/b200-sglang-0.5.14-power-data branch from 6bc4f24 to c8be175 Compare August 26, 2026 14:54

Navigation Menu

Uh oh!

chore(data): add matched B200 SGLang 0.5.14 power - #1535

chore(data): add matched B200 SGLang 0.5.14 power#1535
kaim-eng wants to merge 2 commits into
mainfrom
codex/b200-sglang-0.5.14-power-data

kaim-eng commented Aug 13, 2026 •

edited

Loading

Uh oh!

copy-pr-bot Bot commented Aug 13, 2026

Uh oh!

coderabbitai Bot commented Aug 13, 2026 •

edited

Loading

Review skipped

Uh oh!

github-actions Bot commented Aug 13, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Aug 13, 2026 •

edited

Loading

Uh oh!

tianhaox commented Sep 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

kaim-eng commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Dependency and zero-sentinel update

Summary

Exact-image provenance

Coverage for the packaged stock tables

Overlay policy

Power-quality checks

Rebase and validation

Uh oh!

copy-pr-bot Bot commented Aug 13, 2026

Uh oh!

coderabbitai Bot commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

github-actions Bot commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Perf Parquet Diff Report

Per-File Row Diff Preview

Artifact Contents

Uh oh!

github-actions Bot commented Aug 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Sanity Check Chart Generation Report

Chart Generation Report for system: b200_sxm, backend: sglang, backend_version: 0.5.14

Uh oh!

tianhaox commented Sep 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

kaim-eng commented Aug 13, 2026 •

edited

Loading

coderabbitai Bot commented Aug 13, 2026 •

edited

Loading

github-actions Bot commented Aug 13, 2026 •

edited

Loading

github-actions Bot commented Aug 13, 2026 •

edited

Loading