chore(data): add matched B200 vLLM 0.24.0 power - #1536
Open
kaim-eng wants to merge 2 commits into
Open
Conversation
Contributor
|
Important Review skippedReview was skipped due to path filters ⛔ Files ignored due to path filters (19)
CodeRabbit blocks several paths by default. You can override this behavior by explicitly including those paths in the path filters. For example, including ⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Enterprise Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
Contributor
Perf Parquet Diff Report
Compared
Per-File Row Diff PreviewShowing the first 3 rows per diff kind for each changed parquet file. Full exact CSVs are in aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/vllm/0.24.0/context_attention_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,context_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,bfloat16,0,0,,278.591,,1000.0
VLLM,0.24.0,NVIDIA B200,context_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,bfloat16,0,128,,270.525,,1000.0
VLLM,0.24.0,NVIDIA B200,context_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,fp8,0,0,,272.45,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/attention/vllm/0.24.0/generation_attention_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,generation_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,bfloat16,1,0,,390.036,,1000.0
VLLM,0.24.0,NVIDIA B200,generation_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,bfloat16,1,128,,481.236,,1000.0
VLLM,0.24.0,NVIDIA B200,generation_attention,vllm_flashinfer_trtllmdecode,1,1,1,1,128,1,bfloat16,bfloat16,1023,0,,360.384,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/comm/vllm/0.24.0/custom_allreduce_perf.parquet
modified rows - full 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
vLLM,0.24.0,NVIDIA B200,all_reduce,vLLM_custom_eager,bfloat16,2,1024,vllm_eager,,246.4975,,1000.0
vLLM,0.24.0,NVIDIA B200,all_reduce,vLLM_custom_eager,bfloat16,2,1048576,vllm_eager,,259.6855,,1000.0
vLLM,0.24.0,NVIDIA B200,all_reduce,vLLM_custom_eager,bfloat16,2,128,vllm_eager,,245.9915,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/encoder_attention/vllm/0.24.0/encoder_attention_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,encoder_attention,vllm_vit_flash_attn_fa4,1,1,1,64,bfloat16,,996.018,,1000.0
VLLM,0.24.0,NVIDIA B200,encoder_attention,vllm_vit_flash_attn_fa4,1,1,1,72,bfloat16,,994.989,,1000.0
VLLM,0.24.0,NVIDIA B200,encoder_attention,vllm_vit_flash_attn_fa4,1,1,10,64,bfloat16,,993.674,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/gemm/vllm/0.24.0/gemm_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,gemm_dtype,m,n,k,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,gemm,CutlassFP8ScaledMMLinearKernel,fp8,1,1024,1024,,257.913,,1000.0
VLLM,0.24.0,NVIDIA B200,gemm,CutlassFP8ScaledMMLinearKernel,fp8,1,1024,10240,,474.121,,1000.0
VLLM,0.24.0,NVIDIA B200,gemm,CutlassFP8ScaledMMLinearKernel,fp8,1,1024,12288,,460.3485,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/linear_attention/vllm/0.24.0/gdn_perf.parquet
modified rows - full 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
vLLM,0.24.0,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,16,128,16,128,Qwen/Qwen3.5-0.8B,,242.09,,1000.0
vLLM,0.24.0,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,2,128,2,128,Qwen/Qwen3.5-0.8B,,243.169,,1000.0
vLLM,0.24.0,NVIDIA B200,gdn,causal_conv1d_fn,context,1,1,1,1024,4,4,128,4,128,Qwen/Qwen3.5-0.8B,,247.543,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mhc/vllm/0.24.0/mhc_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,post,vllm.model_executor.kernels.mhc.tilelang.mhc_post_tilelang,DeepseekV4ForCausalLM,1,2,4,4096,20,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,post,vllm.model_executor.kernels.mhc.tilelang.mhc_post_tilelang,DeepseekV4ForCausalLM,1,2,4,7168,20,,243.244,,1000.0
VLLM,0.24.0,NVIDIA B200,post,vllm.model_executor.kernels.mhc.tilelang.mhc_post_tilelang,DeepseekV4ForCausalLM,1024,2,4,4096,20,,247.325,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/vllm/0.24.0/mla_context_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,mla_context_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,1,1,1,0,,866.288,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_context_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,128,1,1,0,,816.4200000000001,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_context_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,16,1,1,0,,847.015,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla/vllm/0.24.0/mla_generation_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,mla_generation_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,1,1,1,1024,,725.5865,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_generation_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,1,1,1,128,,731.839,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_generation_module,FLASHINFER_MLA,deepseek-ai/DeepSeek-V3,DeepseekV3ForCausalLM,bfloat16,bfloat16,bfloat16,1,1,1,1,131072,,402.7006666666667,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/mla_bmm/vllm/0.24.0/mla_bmm_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,bmm_dtype,num_tokens,num_heads,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,mla_gen_post,vllm_torch_bmm,bfloat16,1,1,,266.398,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_gen_post,vllm_torch_bmm,bfloat16,1,12,,255.195,,1000.0
VLLM,0.24.0,NVIDIA B200,mla_gen_post,vllm_torch_bmm,bfloat16,1,128,,260.32,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/moe/vllm/0.24.0/moe_perf.parquet
modified rows - full 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
VLLM,0.1.dev19262+gb6bbf29dd.d20260727,NVIDIA B200,moe,vllm_compressedtensorsw4a4mxfp4moe_marlin_marlinexperts,w4a16_mxfp4,1,3584,3072,16,896,1,1,balanced,,0.0,,0.0
VLLM,0.1.dev19262+gb6bbf29dd.d20260727,NVIDIA B200,moe,vllm_compressedtensorsw4a4mxfp4moe_marlin_marlinexperts,w4a16_mxfp4,1,3584,3072,16,896,1,1,power_law_1.01,,0.0,,0.0
VLLM,0.1.dev19262+gb6bbf29dd.d20260727,NVIDIA B200,moe,vllm_compressedtensorsw4a4mxfp4moe_marlin_marlinexperts,w4a16_mxfp4,1,3584,3072,16,896,1,1,power_law_1.2,,0.0,,0.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/vllm/0.24.0/computescale_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,m,k,quant_dtype,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,compute_scale,dynamic_per_token_scaled_fp8_quant_minus_static_scaled_fp8_quant,1,1024,fp8,,259.451,,1000.0
VLLM,0.24.0,NVIDIA B200,compute_scale,dynamic_per_token_scaled_fp8_quant_minus_static_scaled_fp8_quant,1,10240,fp8,,258.907,,1000.0
VLLM,0.24.0,NVIDIA B200,compute_scale,dynamic_per_token_scaled_fp8_quant_minus_static_scaled_fp8_quant,1,12288,fp8,,276.849,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/quantize/vllm/0.24.0/scale_matrix_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,m,k,quant_dtype,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,scale_matrix,static_scaled_fp8_quant,1,1024,fp8,,272.792,,1000.0
VLLM,0.24.0,NVIDIA B200,scale_matrix,static_scaled_fp8_quant,1,10240,fp8,,258.902,,1000.0
VLLM,0.24.0,NVIDIA B200,scale_matrix,static_scaled_fp8_quant,1,12288,fp8,,276.849,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_csa_context_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,dsv4_csa_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,0,4,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_csa_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,128,4,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_csa_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,2048,4,,0.0,,0.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_csa_generation_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,dsv4_csa_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,1,4,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_csa_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,1024,4,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_csa_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,10240,4,,0.0,,0.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_hca_attn_module_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,local_num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,dsv4_hca_attn_module,FLASHMLA_SPARSE_DSV4,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,64,64,1,1,1,0,128,,259.925,,1000.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_attn_module,FLASHMLA_SPARSE_DSV4,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,64,64,1,1,1,1,128,,259.233,,1000.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_attn_module,FLASHMLA_SPARSE_DSV4,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,64,64,1,1,1,1024,128,,259.467,,1000.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_hca_context_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,dsv4_hca_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,0,128,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,128,128,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_context_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,2048,128,,0.0,,0.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_hca_generation_module_perf.parquet
modified rows - full 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
VLLM,0.24.0,NVIDIA B200,dsv4_hca_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,1,128,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,1024,128,,0.0,,0.0
VLLM,0.24.0,NVIDIA B200,dsv4_hca_generation_module,FLASHMLA_SPARSE_DSV4,deepseek-ai/DeepSeek-V4-Pro,DeepseekV4ForCausalLM,bfloat16,fp8,fp8_block,128,1,1,1,10240,128,,0.0,,0.0aic-core/src/aiconfigurator_core/systems/data/b200_sxm/sparse_attention/vllm/0.24.0/dsv4_paged_mqa_logits_module_perf.parquet
modified rows - full CSV: framework,version,device,op_name,kernel_source,model,architecture,mla_dtype,kv_cache_dtype,gemm_type,num_heads,local_num_heads,batch_size,isl,tp_size,step,compress_ratio,power__base,power__head,power_limit__base,power_limit__head
VLLM,0.24.0,NVIDIA B200,dsv4_paged_mqa_logits_module,vllm.utils.deep_gemm.fp8_fp4_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8,fp8_block,64,64,1,1,1,0,4,,246.992,,1000.0
VLLM,0.24.0,NVIDIA B200,dsv4_paged_mqa_logits_module,vllm.utils.deep_gemm.fp8_fp4_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8,fp8_block,64,64,1,1,1,1,4,,250.742,,1000.0
VLLM,0.24.0,NVIDIA B200,dsv4_paged_mqa_logits_module,vllm.utils.deep_gemm.fp8_fp4_paged_mqa_logits,sgl-project/DeepSeek-V4-Flash-FP8,DeepseekV4ForCausalLM,fp8_e4m3,fp8,fp8_block,64,64,1,1,1,1024,4,,247.1,,1000.0Artifact Contents
|
Contributor
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 Below is a report of whether the chart generation was successful for each op. Chart Generation Report for system: b200_sxm, backend: vllm, backend_version: 0.24.0
|
kaim-eng
changed the title
kaim-eng
force-pushed
the
codex/b200-vllm-0.24.0-power-data
branch
2 times, most recently
from
August 13, 2026 16:59
32f0781 to
5960346
Compare
kaim-eng
force-pushed
the
codex/b200-vllm-0.24.0-power-data
branch
from
August 21, 2026 13:52
ab1429f to
dcc4320
Compare
kaim-eng
changed the base branch from
codex/b200-power-collector-fixes
to
main
August 21, 2026 13:53
kaim-eng
force-pushed
the
codex/b200-vllm-0.24.0-power-data
branch
from
August 24, 2026 12:59
34d8bce to
e384fc5
Compare
kaim-eng
marked this pull request as ready for review
August 24, 2026 14:27
Signed-off-by: Kai Ma <kaim@nvidia.com>
Signed-off-by: kaim-eng <kaim@NV-8QHBYK4.localdomain>
kaim-eng
force-pushed
the
codex/b200-vllm-0.24.0-power-data
branch
from
August 26, 2026 14:54
e384fc5 to
dc7597a
Compare
Contributor
|
Independent Arrow audit of the 19 files at dc7597a against the merge-base (2ed278a): all 375,513 rows retained, every non-power column value- and order-identical to stock, The only conflict with current |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Important
Dependency and zero-sentinel update
This pure-data PR depends on #1590. Merge #1590 first, then rebase/update this PR onto the resulting
mainbefore merging it.Latest signed-off data head:
e384fc5e50d7c6bf0cc37b6a48231e161994e2cf.The latest commit normalizes every unavailable
power/power_limitcell in the 19 changed power-enabled tables to typed0.0. Row coverage, measured values, logical keys, and physical row order are unchanged; the two unchanged DSA tables remain power-disabled and column-free.Summary
Pure-data PR constructed from the then-current
maindata snapshot (095f58a51c4ca8e61b66ec108d86f223f8d559ce), refreshed from the completed B200 vLLM power campaign.vllm/vllm-openai:v0.24.0@sha256:251eba5cc7c12fed0b75da22a9240e582b1c9e39f6fbc064f86781b963bd814f0c6d722ecbbbce39c477adceae2687eb7245d22cfloat64powerandpower_limiton exact logical-key matchesFinal coverage for the packaged stock tables
Across all 21 current-main B200 vLLM 0.24.0 tables:
0.0pairs in changed tables plus 25,639 rows in the unchanged power-disabled DSA tables)power/power_limitpairs: 0The PR changes 19 tables containing 375,513 rows; 361,348 are powered and 14,165 use typed
0.0sentinel pairs. Eighteen files are byte-identical to the prior PR head. The stock MoE table gained 2,025 rows after the collection snapshot; those rows are preserved with0.0sentinel pairs because no exact collected key exists. The two DSA tables remain unchanged, power-disabled, and column-free (25,639 rows).Of the powered rows:
Raw row coverage is not a substitute for the estimator's existing fail-closed, latency-weighted public-estimate coverage gate.
Latest artifact reconciliation
The finalized campaign output contains 414,432 raw rows. After applying the Collector identity rule (all columns except
latency,power, andpower_limit):Collected-only keys by family:
Promotion policy
A candidate table must have median absolute latency drift at or below 25% against frozen stock. Within promoted tables, each row must also have:
Stock latency is always retained; only power fields are overlaid.
The latest run newly promotes:
It also adds newly qualifying stock rows to prior tables:
Custom all-reduce had no candidate in this campaign and retains its 138 previously reviewed power rows.
DSA disposition
The collection produced exact matches for all 14,640 DSA-context stock rows and 10,995 of 10,999 DSA-generation stock rows. Those counts demonstrate identity coverage, not promotion validity.
Both tables remain byte-identical and perf-only because their median absolute latency drift still exceeds the agreed table gate:
The 41 DSA-generation collected-only keys are excluded. The four unmatched stock generation keys remain unchanged.
Runtime safety and merge gate
This PR contains no code changes and depends on #1590 for the zero-sentinel loader contract. Targeted support-matrix validation with #1590 produced 122
PASS, 2HYBRID_PASS, and 18 unrelatedFAILrows, with zero null-access errors.Validation
maindata snapshot095f58a51c4ca8e61b66ec108d86f223f8d559ce; rebase onto the post-fix(collector): normalize missing power metrics to zero #1590mainis required before mergee384fc5e50d7c6bf0cc37b6a48231e161994e2cf0.0sentinel pairs, 0 null or partial pairsencoder_attention/vllm/0.24.0/encoder_attention_perf.parquetThe legacy
collection_meta.yamlfiles are intentionally untouched. This overlay combines reviewed measurements from the completed campaign output, while this PR records the exact source image, collector commit, exclusions, and promotion policy.