perf-changelog.yaml

- config-keys:
    - dsr1-fp8-b200-dynamo-trt
    - dsr1-fp8-h200-dynamo-trt
    - dsr1-fp4-gb200-dynamo-trt
  description:
    - "Fix metadata inconsistencies in nvidia-master.yaml - TP/EP/DP-attn values now match actual recipe files"
    - "B200 FP8 TRT 8K/1K: prefill_ep 8→1 (15 entries), prefill_dp_attn true→false (1 entry)"
    - "H200 FP8 TRT 1K/1K: prefill_dp_attn false→true (9 entries)"
    - "H200 FP8 TRT 8K/1K: prefill_dp_attn true→false (8 entries)"
    - "GB200 FP4 TRT 8K/1K: decode_dp_attn true→false (2 entries)"
    - "All fixes are metadata-only; no recipe files were modified"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/919

- config-keys:
    - kimik2.5-int4-mi325x-vllm
  description:
    - "Add Kimi K2.5 INT4 single-node MI325X vLLM benchmark (TP8)"
    - "Uses vLLM ROCm v0.16.0 image following AMD Andy Luo's recipe"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/857

- config-keys:
    - 70b-fp8-*-vllm
  description:
    - 'Add compilation-config ''{"custom_ops": ["-rms_norm", "-quant_fp8", "-silu_and_mul"]}'' as extra config to all benchmarks/70b_fp8_mi*.sh scripts'
    - "6-7% uplift for llama for 6/8 configs"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/95

- config-keys:
    - gptoss-fp4-*-trt
  description:
    - "Upgrade GPT-OSS TRT images from 'release:1.1.0rc2.post2' to '1.2.0rc0.post1'"
    - "Add NCCL_GRAPH_REGISTER=0 to benchmarks/gptoss_fp4_b200_trt_slurm.sh"
    - "Change kv_cache_config.dtype from 'auto' to 'fp8' in benchmarks/gptoss_fp4_b200_trt_slurm.sh"
    - "Remove MOE_BACKEND=CUTLASS, now just defaults to TRTLLM"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/110

- config-keys:
    - gptoss*
    - dsr1*
  description:
    - "Remove Llama 70B runs to make room for multi-node disagg prefill+wideEP on h100/h200/b200/mi300/mi325/mi355"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/149

- config-keys:
    - gptoss-fp4-b200-vllm
    - gptoss-fp4-h100-vllm
    - gptoss-fp4-h200-vllm
  description:
    - "Upgrade vLLM from 0.10.2 to 0.11.0 for GPT-OSS NVIDIA single-node configs"
    - 'Add compilation-config ''{"cudagraph_mode":"PIECEWISE"}'' since vLLM 0.11.0 now defaults to FULL_AND_PIECEWISE'
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/159

- config-keys:
    - dsr1*
  description:
    - "Fix bug where 1k8k and 8k1k full sweeps had incorrect max-model-len for DeepSeek"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/163

- config-keys:
    - dsr1-fp4-b200-sglang
    - dsr1-fp8-b200-sglang
    - dsr1-fp8-h200-sglang
  description:
    - "Consolidate H200 and B200 SGLang configurations to use unified v0.5.5-cu129-amd64 image tag"
    - "Update deprecated SGLang server arguments to current equivalents"
    - "Replace --enable-ep-moe with --ep-size $EP_SIZE"
    - "Replace --enable-flashinfer-trtllm-moe with --moe-runner-backend flashinfer_trtllm"
    - "Add -e EP_SIZE to Docker run commands in launch scripts"
    - "Set ep:4 for all tp:4 entries, ep:8 for all tp:8 entries"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/204

- config-keys:
    - gptoss-fp4-mi355x-vllm
    - gptoss-fp4-b200-vllm
  description:
    - "Extend concurrency to 128 for gptoss mi355x/b200 vllm configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/209

- config-keys:
    - gptoss-fp4-b200-trt
  description:
    - "Extend concurrency to 128 for gptoss b200 TRT configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/233

- config-keys:
    - "*gb200-dynamo-sglang"
  description:
    - "Introduce improvements in GB200 SGLang DSR1 submission"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/257

- config-keys:
    - dsr1-fp8-h200-trt
  description:
    - "Update TRT image from nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc0.post1 to nvcr.io#nvidia/tensorrt-llm/release:1.2.0rc2"
    - "Increase concurrency for some configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/266

- config-keys:
    - gptoss-fp4-b200-vllm
    - gptoss-fp4-h100-vllm
    - gptoss-fp4-h200-vllm
  description:
    - "Update vLLM image for NVIDIA configs from vLLM 0.11.0 to vLLM 0.11.2"
    - "Add kv-cache-dtype: fp8 to benchmarks/gptoss_fp4_b200_docker.sh"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/273

- config-keys:
    - gptoss-fp4-b200-trt
  description:
    - "Add benchmark script for GPTOSS FP4 B200 TRT-LLM"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/256

- config-keys:
    - dsr1-fp4-gb200-dynamo-trt
    - dsr1-fp4-gb200-dynamo-sglang
    - dsr1-fp8-gb200-dynamo-sglang
  description:
    - "Add more configurations for GB200 SGLang DSR1"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/335

- config-keys:
    - dsr1-fp4-mi355x-sglang
  description:
    - "Update MI355x Deepseek-R1 FP4 SGLang Image to upstream v0.5.6.post1"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/330

- config-keys:
    - dsr1-fp4-gb200-dynamo-sglang
    - dsr1-fp8-gb200-dynamo-sglang
  description:
    - "fix: Pruning unnecessary concurrencies "
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/358

- config-keys:
    - dsr1-fp4-mi355x-sglang
  description:
    - "Updating MI355x Deepseek-R1 FP4 SGLang Image to upstream v0.5.6.post2"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/369
    
- config-keys:
    - dsr1-fp4-b200-sglang
    - dsr1-fp8-b200-sglang
    - dsr1-fp8-h200-sglang
  description:
    - "Update NVIDIA DeepSeek sglang Docker image from v0.5.5 to v0.5.6"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/276
  
- config-keys:
    - gptoss-fp4-b200-vllm
    - gptoss-fp4-h100-vllm
    - gptoss-fp4-h200-vllm
  description: 
    - "Update vLLM image from v0.11.2 to v0.13.0"
    - "Add VLLM_MXFP4_USE_MARLIN=1 to H100 and H200 benchmark scripts"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/327

- config-keys:
    - dsr1-fp8-mi300x-sglang
    - dsr1-fp8-mi325x-sglang
    - dsr1-fp8-mi355x-sglang
  description:
    - Use upstream SGLang images on mi300, mi325 and mi355 for dsr1fp8
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/332

- config-keys:
    - gptoss-fp4-gb200-dynamo-trt
    - gptoss-fp4-b200-trt
  description:
    - Explicitly add EP=TP for DP attention configs for B200 AGG nvidia-master file. Multinode Refactor inadvertently changed default EP=1
    - Add GPTOSS DISAGG configurations for GB200 1k1k and 8k1k.
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/387

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Add PD disaggregation (1P2D) for Mi355X"
    - "Includes with and without speculative decoding"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/348

- config-keys:
    - dsr1-fp4-mi355x-sglang
  description:
    - "Updating MI355x Deepseek-R1 FP4 SGLang Image to upstream v0.5.7"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/395

- config-keys:
    - dsr1-fp8-b200-sglang
  description:
    - "Adds TP4 configurations to DSR1-FP8 B200 SGLang deployment experiments"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/411
  
- config-keys:
    - dsr1-fp4-b200-trt-mtp
    - dsr1-fp8-b200-trt-mtp
    - dsr1-fp8-h200-trt-mtp
  description:
    - Add MTP (Multi-Token Prediction) support for single-node TRT configs
    - Add spec-decoding field to config entries and update launch scripts to select MTP benchmark scripts
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/392
  
- config-keys:
    - dsr1-fp8-mi355x-atom
    - dsr1-fp4-mi355x-atom
    - gptoss-fp4-mi355x-atom
  description:
    - Add internal AMD ATOM inference engine for DeepSeek R1 FP8, FP4 and GPTOSS FP4 Mi355X
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/419

- config-keys:
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
  description:
    - "Update AMD MI300X and MI325X GPT-OSS 120B vLLM to use upstream ROCm image vllm/vllm-openai-rocm:v0.14.0"
    - "Remove deprecated --async-scheduling flag (now enabled by default in vLLM v0.14.0)"
    - "Remove deprecated --max-seq-len-to-capture flag"
    - "Add HIP_VISIBLE_DEVICES env var for Ray compatibility in vLLM 0.14+"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/496

- config-keys:
    - dsr1-fp8-h200-sglang
  description:
    - "Update H200 DeepSeek R1 FP8 SGLang image from v0.5.6 to v0.5.7"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/538

- config-keys:
    - dsr1-fp8-mi300x-sglang
  description:
    - "Update MI300X DeepSeek R1 FP8 SGLang image from v0.5.5.post3 to v0.5.7"
    - "Add SGLANG_AITER_MLA_PERSIST=1 for persistent MLA kernel optimization"
    - "Set --kv-cache-dtype fp8_e4m3 for fp8 KV cache"
    - "Set --attention-backend aiter for AMD aiter attention backend"
    - "Update chunked-prefill-size and max-prefill-tokens from 196608 to 131072"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/544
- config-keys:
    - dsr1-fp8-mi325x-sglang
  description:
    - "Update MI325X DeepSeek R1 FP8 SGLang image from v0.5.5.post3 to v0.5.7"
    - "Add SGLANG_AITER_MLA_PERSIST=1 for persistent MLA kernel with fp8 KV cache"
    - "Add --kv-cache-dtype fp8_e4m3 for explicit FP8 KV cache"
    - "Add --attention-backend aiter for AMD aiter attention backend"
    - "Reduce chunked-prefill-size from 196608 to 131072"
    - "Reduce max-prefill-tokens from 196608 to 131072"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/545

- config-keys:
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
  description:
    - "Fix AITER env vars for vLLM v0.14.0 on AMD MI300X and MI325X"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/535

- config-keys:
    - dsr1-fp8-mi355x-sglang
  description:
    - "Update MI355X DeepSeek R1 FP8 SGLang image from v0.5.5.post3 to v0.5.8"
    - "Key fix: Disables mla persistent kernel when not using fp8 kv_cache (https://github.com/sgl-project/sglang/pull/17327)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/572

- config-keys:
    # NVIDIA single-node
    - dsr1-fp4-b200-sglang
    - dsr1-fp4-b200-trt
    - dsr1-fp4-b200-trt-mtp
    - dsr1-fp8-b200-sglang
    - dsr1-fp8-b200-trt
    - dsr1-fp8-b200-trt-mtp
    - dsr1-fp8-h200-sglang
    - dsr1-fp8-h200-trt
    - dsr1-fp8-h200-trt-mtp
    - gptoss-fp4-b200-trt
    - gptoss-fp4-b200-vllm
    - gptoss-fp4-h100-vllm
    - gptoss-fp4-h200-trt
    - gptoss-fp4-h200-vllm
    # AMD single-node
    - dsr1-fp4-mi355x-sglang
    - dsr1-fp4-mi355x-atom
    - dsr1-fp8-mi300x-sglang
    - dsr1-fp8-mi325x-sglang
    - dsr1-fp8-mi355x-sglang
    - dsr1-fp8-mi355x-atom
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
    - gptoss-fp4-mi355x-vllm
    - gptoss-fp4-mi355x-atom
  description:
    - Add official GSM8k eval results to GPT-OSS and DeepSeek R1 scenarios
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/558
  evals-only: true

- config-keys:
    - dsr1-fp8-h200-sglang
  description:
    - "Update H200 DeepSeek R1 FP8 SGLang image from v0.5.7 to v0.5.9"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
  
- config-keys:
  - dsr1-fp4-b300-dynamo-trt
  description:
    - "Add DSR1 FP4 B300 Dynamo TRT configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/585

- config-keys:
  - dsr1-fp4-mi355x-sglang
  description:
    - "Update SGLang image from v0.5.7 to v0.5.8 for DeepSeek-R1 FP4 on MI355x"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/595
  
- config-keys:
    - dsr1-fp8-b200-trt
  description:
    - "Update TensorRT-LLM container from release:1.1.0rc2.post2 to release:1.2.0rc6.post2"
    - "Change default MOE backend from DEEPGEMM to TRTLLM"
    - "Add dynamic piecewise CUDA graphs for 1k1k (CONC≥64) and 8k1k (CONC≥64) workloads"
    - "Add delay batching (batch_wait_timeout_iters/batch_wait_max_tokens_ratio) for 1k1k high-concurrency"
    - "Add dynamic KV cache memory fraction tuning (0.7-0.8) based on ISL/OSL/TP configuration"
    - "Update search space: remove EP=TP constraint, add TP=4 configurations, extend concurrency ranges"
    - "Add TLLM_OVERRIDE_LAYER_NUM=61 to avoid OOM errors"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/594
  
- config-keys:
  - dsr1-fp4-b200-dynamo-trt
  description:
    - "Update DSR1 FP4 B200 Dynamo TRT configurations"
    - "Update TRTLLM version to 1.2.0rc6.post2"
    - "Transform to use srt-slurm recipes"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/588

- config-keys:
    - dsr1-fp8-h200-dynamo-trt
  description:
    - "Add DSR1 FP8 H200 Dynamo TRT-LLM disaggregated multinode configuration"
    - "Image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.8.1.post1"
    - "Runner: h200-dgxc with multinode and disagg enabled"
    - "Includes MTP and STP configurations for 1k1k and 8k1k sequence lengths"
    - "Concurrency levels: 4, 8, 16, 32, 64, 128, 256, 512"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/570

- config-keys:
    - dsr1-fp4-gb200-dynamo-trt
  description:
    - "Update Dynamo TRT image from 0.5.1-rc0.pre3 to 0.8.1.post2"
    - "Update TRT configurations"
    - "Refactor configurations to use CONFIG_FILE-based recipes instead of inline parameter settings"
    - "Introduce srt-slurm workflow for launching Dynamo jobs"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/510

- config-keys:
    - dsr1-fp8-mi355x-sglang
  description:
    - "Disable torch.compile for MI355X DeepSeek-R1 FP8 SGLang"
    - "set cuda-graph-max-bs to CONC"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/613

- config-keys:
    - dsr1-fp8-h200-dynamo-sglang
  description:
    - "Add DSR1 FP8 H200 Dynamo SGLang disaggregated multinode configuration"
    - "Image: lmsysorg/sglang:v0.5.8-cu130-runtime"
    - "Runner: h200-multinode-slurm with multinode and disagg enabled"
    - "Recipes sourced from srtslurm repo (recipes/h200/)"
    - "1k1k configs: aggregated, low-latency (1P9D), high-throughput TEP (1P6D), DEP (1P6D)"
    - "8k1k configs: aggregated, TEP configs (1P7D, 1P6D, 1P3D, 2P3D), DEP (1P1D)"
    - "Concurrency levels range from 1 to 2048 depending on configuration"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/582

- config-keys:
    - dsr1-fp4-b200-trt
  description:
    - "Update TensorRT-LLM container from release:1.1.0rc2.post2 to release:1.2.0rc6.post2"
    - "Change default MOE backend from DEEPGEMM to TRTLLM"
    - "Add dynamic piecewise CUDA graphs for 1k1k (TEP8 and CONC64)"
    - "Update search space: remove EP=TP constraint, add TP=4 configurations, extend concurrency ranges"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/620

- config-keys:
    - dsr1-fp4-gb300-dynamo-trt
  description:
    - "Add DeepSeek-R1 FP4 GB300 Dynamo TRT disaggregated multinode configurations"
    - "Image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.8.1.post2"
    - "Includes MTP and STP configs for 1k1k, 1k8k, and 8k1k sequence lengths"
    - "Add gb300-nv runner and launch script for srt-slurm integration"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/618

- config-keys:
    - dsr1-fp4-mi355x-sglang-disagg
  description:
    - "enable PD/D for both MTP and non-MTP MI355X DeepSeek-R1 FP4 SGLang"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/622

- config-keys:
    - dsr1-fp8-gb200-dynamo-trt
  description:
    - "Add DeepSeek R1 FP8 GB200 Dynamo TRT-LLM disaggregated multinode configurations"
    - "Image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:0.8.1.post2"
    - "1k1k: 14 scenarios (7 MTP, 7 STP) with varying DP attention/TEP modes"
    - "1k8k: 10 scenarios (5 MTP, 5 STP) for long output generation"
    - "8k1k: 14 scenarios (7 MTP, 7 STP) for long context workloads"
    - "Prefill workers: 1-5P, Decode workers: 1-4D, TP/EP: 8/16/32"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/617

- config-keys:
    - dsr1-fp8-gb200-dynamo-trt
  description:
    - "Fix model_prefix argument in yaml configs"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/646

- config-keys:
    - dsr1-fp8-b200-trt-mtp
  description:
    - Update to the latest TRTLLM 1.2 release container version
    - Fine-tune the choice of parallelism in nvidia-master file, mainly going to TP for most points
    - Enable piecewise CUDA graphs under most conditions
    - fine-tune max batch sizes and other optimizations
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/632

- config-keys:
    - dsr1-fp8-gb300-dynamo-trt
  description:
    - "Add DeepSeek R1 FP8 GB300 Dynamo TRT-LLM disaggregated multinode configurations for 8k1k and 1k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/627

- config-keys:
    - gptoss-fp4-b200-trt
  description:
    - "Update GPT-OSS FP4 B200 TRT pareto configurations and new container image"
    - "Extend maximum concurrency to 256 across all sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/639

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Add --use-chat-template argument to benchmark_serving script"
    - "Without this arg, MTP acceptance rates are artificially high for DeepSeek with MTP"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/647

- config-keys: 
    - dsr1-fp8-b200-sglang-mtp
  description:
    - "Add MTP (Multi-Token Prediction) support for DeepSeek R1 FP8 B200 SGLang using EAGLE speculative decoding"
    - "Image: lmsysorg/sglang:v0.5.8-cu130-amd64"
    - "Add benchmark script dsr1_fp8_b200_mtp.sh with EAGLE speculative decoding (num-steps=2, draft-tokens=3, topk=1)"
    - "Update launch_b200-dgxc.sh to support SPEC_SUFFIX for MTP script selection"
    - "Configurations: TP=8, EP=1, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/626

- config-keys:
    - dsr1-fp4-b200-trt-mtp
  description:
    - "Upgrade TensorRT-LLM container from release:1.1.0rc2.post2 to release:1.2.0rc6.post3"
    - "Enable dynamic piecewise CUDA graphs for several conditions"
    - "Adjust TP8/TP4 search space to reduce overlapping points"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/642

- config-keys:
    - dsr1-fp8-b200-dynamo-sglang
  description:
    - "Add DSR1 FP8 B200 disaggregated SGLang multinode configuration"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-amd64"
    - "9 recipes: 4x 1k1k + 5x 8k1k, low-latency and max-throughput profiles"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/658

- config-keys:
    - dsr1-fp8-gb300-dynamo-trt
  description:
    - "Add DeepSeek R1 FP8 GB300 Dynamo TRT-LLM disaggregated multinode configurations for 1k8k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/654

- config-keys:
    - dsr1-fp8-h100-dynamo-trt
  description:
    - "Add DeepSeek R1 FP8 H100 Dynamo TRT-LLM disaggregated multinode configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/651

- config-keys:
    - dsr1-fp8-h200-dynamo-sglang
  description:
    - "Add MTP (EAGLE speculative decoding) configs alongside STP"
    - "Update container to lmsysorg/sglang:v0.5.8.post1-cu130"
    - "Remove aggregated configs, keep only disaggregated"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/640

- config-keys:
    - dsr1-fp8-b300-dynamo-trt
  description:
    - "New B300 FP8 Dynamo TRT configurations"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/638

- config-keys:
    - dsr1-fp8-h100-dynamo-trt
  description:
    - "Add DeepSeek R1 FP8 H100 Dynamo TRT-LLM disaggregated multinode configurations"
    - "fix model_prefix bug from https://github.com/SemiAnalysisAI/InferenceX/pull/651"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/663

- config-keys:
    - dsr1-fp4-gb200-dynamo-sglang
  description:
    - "Update SGLang image from v0.5.5.post2 to v0.5.8-cu130"
    - "Add FP4 model path separation via SRT_SLURM_MODEL_PREFIX in launch script"
    - "Refactor to use CONFIG_FILE-based srt-slurm recipes instead of inline parameters"
    - "Add 1k1k configurations: low-latency (1P2D), mid-curve (4P8D), max-tpt (4P12D)"
    - "Add 8k1k configurations: low-latency (1P4D), mid-curve (6P12D), max-tpt (10P8D)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/633  

- config-keys:
    - dsr1-fp8-gb200-dynamo-sglang
    - dsr1-fp8-gb300-dynamo-sglang
  description:
    - "Update  GB200 and GB300 configs for DSR1 FP8 SGLANG STP mode"
    - "Image: lmsysorg/sglang:v0.5.8-cu130"
    - "Update prefill/decode worker counts, TP/EP parallelism, and dp-attn settings for 1k1k and 8k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/635

- config-keys:
    - dsr1-fp8-b200-dynamo-sglang-mtp
  description:
    - "Add DSR1 FP8 B200 disaggregated SGLang MTP multinode configuration"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-amd64"
    - "9 recipes: 4x 1k1k + 5x 8k1k, low-latency and max-throughput with EAGLE speculative decoding"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/667

- config-keys:
    - dsr1-fp8-h100-dynamo-sglang
  description:
    - "Add DeepSeek-R1 FP8 H100 Dynamo SGLang STP disaggregated multinode configurations"
    - "Image: lmsysorg/sglang:v0.5.8-cu130"
    - "1k1k, 1k8k, 8k1k sequence lengths"
    - "Two modes per seq-len: Max throughput TEP (1P2D) and Max throughput DEP (1P1D with dp-attention)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/643

- config-keys:
    - dsr1-fp8-h100-dynamo-sglang
  description:
    - "Add DeepSeek-R1 FP8 H100 Dynamo SGLang MTP disaggregated multinode configurations"
    - "Image: lmsysorg/sglang:v0.5.8-cu130"
    - "1k1k, 1k8k, 8k1k sequence lengths with MTP speculative decoding"
    - "Two modes per seq-len: Max throughput TEP (1P2D) and Max throughput DEP (1P1D with dp-attention)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/644

- config-keys:
    - dsr1-fp8-mi355x-atom-mtp
    - dsr1-fp4-mi355x-atom-mtp
  description:
    - "Add DSR1 FP8/FP4 MI355X ATOM with MTP configuration"
    - "Image: rocm/atom:rocm7.2.0-ubuntu24.04-pytorch2.9-atom0.1.1"
    - "Deepseek R1 with speculative decoding: 1k1k, 1k8k, 8k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/673

- config-keys:
    - dsr1-fp4-b200-dynamo-sglang
  description:
    - "Add DSR1 FP4 B200 Dynamo SGLang STP mode"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-runtime"
    - "1k1k configs: low-latency DEP (1P5D, 1P6D), max-throughput DEP (1P1D, 1P2D)"
    - "8k1k configs: low-latency DEP/TEP (1P1D, 1P5D, 2P5D), TEP (1P1D), max-throughput DEP (7P2D)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/672

- config-keys:
    - dsr1-fp8-b200-dynamo-trt
  description:
    - "Introduce new DSR1 FP8 B200 Dynamo TRT configurations for 8k1k and 1k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/616

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
    - dsr1-fp4-mi355x-sglang-disagg
  description:
    - "Bump MI355X MORI FP8/FP4 image to latest rocm/sgl-dev:sglang-0.5.8-rocm700-mi35x-mori-0210"
    - "Bump mi355x sglang disagg recipe to sa-260211"
    - "Add conc 4/8/16"
    - "Use Pure TP with MTP=2 for 1k1k conc smaller than 128 and reduce MTP to 1 for DEP configs"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/674

- config-keys:
    - dsr1-fp4-b200-dynamo-sglang-mtp
  description:
    - "Add B200 configs for DSR1 FP4 SGLANG MTP mode for 1k1k and 8k1k"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/683

- config-keys:
    - dsr1-fp4-gb300-dynamo-sglang
  description:
    - "Add GB300 FP4 Dynamo SGLang disaggregated multinode configuration"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-runtime"
    - "Recipes sourced from srt-slurm repo (recipes/gb300-fp4/ folder)"
    - "Add 1k1k configurations: low-latency (1P2D), mid-curve (4P8D), max-tpt (4P12D)"
    - "Add 8k1k configurations: low-latency (1P4D), mid-curve (6P12D), max-tpt (10P8D)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/636

- config-keys:
    - dsr1-fp8-b200-dynamo-sglang-mtp
  description:
    - "Patches one missing concurrency point for "
    - "DSR1 FP8 B200 disaggregated SGLang MTP multinode configuration. "
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-amd64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/691

- config-keys:
    - dsr1-fp8-b300-dynamo-trt
  description:
    - "Update max_num_tokens and max_batch_size for min-latency decode workers"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/690

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Add more sweep points for DSR1 FP8 both MTP and non-MTP 1k1k, 8k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/689

- config-keys:
    - dsr1-fp8-b200-dynamo-trt
  description:
    - "Update max_num_tokens and max_batch_size for min-latency decode workers"
    - "See srt-slurm recipe changes: https://github.com/ishandhanani/srt-slurm/pull/173"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/686

- config-keys:
    - dsr1-fp8-mi325x-sglang
  description:
    - "Update MI325X DeepSeek R1 FP8 SGLang image from v0.5.7 to v0.5.8"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/692

- config-keys:
    - dsr1-fp8-mi300x-sglang
  description:
    - "Update MI300X DeepSeek R1 FP8 SGLang image from v0.5.7 to v0.5.8"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/696

- config-keys:
    - dsr1-fp4-mi355x-sglang-disagg
  description:
    - "Add more sweep points for DSR1 FP4 both MTP and non-MTP 1k1k, 8k1k"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/692
  
- config-keys:
    - dsr1-fp8-b200-dynamo-sglang-mtp
  description:
    - "Add new 1P2D max-throughput MTP config for 1k1k"
    - "MTP settings: speculative-num-steps=1, speculative-num-draft-tokens=2"
    - "Image: lmsysorg/sglang:v0.5.8.post1-cu130-amd64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/697

- config-keys:
    - dsr1-fp8-h200-dynamo-trt
  description:
    - "Add min-latency configurations for H200 Dynamo TRT"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/698

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Bump MI355X disagg FP8 recipe commit to 953f7c5 (bug fixes in sglang_disagg fork)"
    - "1k1k: Switch prefill workers from DEP (tp:1, ep:8, dp-attn) to Pure TP (tp:8, ep:1) for MTP and non-MTP middle-of-curve 1P2D configs"
    - "1k1k: Extend middle-of-curve concurrency range by adding conc=128 for both MTP and non-MTP 1P2D configs"
    - "8k1k: Add new 2P1D (2-prefill, 1-decode) configs at conc [512, 1024] for both MTP (DECODE_MTP_SIZE=0) and non-MTP, with Pure TP prefill and DEP decode"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/700

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Bump MI355X disagg FP8 recipe commit to fix perf regression on 8k1k DEP8"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/701

- config-keys:
    - qwen3.5-bf16-b200-sglang
  description:
    - "Add Qwen3.5-397B-A17B BF16 B200 SGLang benchmark"
    - "Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/704

- config-keys:
    - qwen3.5-bf16-mi355x-sglang
  description:
    - "Add Qwen3.5-397B-A17B BF16 SGLang benchmark for MI355X"
    - "Image: rocm/sgl-dev:v0.5.8.post1-rocm720-mi35x-20260215"
    - "Uses triton attention backend, TP=8, concurrency 4-64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/705

- config-keys:
    - kimik2.5-int4-mi355x-vllm
  description:
    - "Add Kimi-K2.5 INT4 vLLM benchmark for MI355X"
    - "Model: moonshotai/Kimi-K2.5 with --mm-encoder-tp-mode data"
    - "Image: vllm/vllm-openai-rocm:v0.15.1"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/734

- config-keys:
    - minimaxm2.5-fp8-mi355x-vllm
  description:
    - "Add MiniMax-M2.5 FP8 vLLM benchmark for MI355X"
    - "Model: MiniMaxAI/MiniMax-M2.5 with --trust-remote-code"
    - "Image: vllm/vllm-openai-rocm:v0.15.1"
    - "Environment: VLLM_ROCM_USE_AITER=1"
    - "TP=2 and TP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/755

- config-keys:
    - qwen3.5-fp8-mi355x-sglang
  description:
    - "Add Qwen3.5-397B-A17B-FP8 SGLang benchmark configuration for MI355X"
    - "Image: rocm/sgl-dev:v0.5.8.post1-rocm720-mi35x-20260218"
    - "Uses triton attention backend, TP=8, concurrency 4-64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/768

- config-keys:
    - qwen3.5-bf16-b200-sglang
  description:
    - "Update Qwen3.5-397B-A17B BF16 SGLang B200 benchmark launch config"
    - "Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e"
    - "Add trtllm_mha attention backend, flashinfer_trtllm MOE runner"
    - "Add context-length, tokenizer-worker-num, env tuning (NCCL_NVLS_ENABLE, SGLANG_ENABLE_FLASHINFER_GEMM)"
    - "Set cuda-graph-max-bs to match concurrency, scheduler-recv-interval based on concurrency"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/758

- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
  description:
    - "Add more configs for MI355X FP8 Disagg"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/770
  
- config-keys:
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
  description:
    - "Update vLLM ROCm image from v0.14.0 to v0.15.1 for MI300X and MI325X GPT-OSS"
    - "Gains: ROCm skinny GEMM dispatch fix, MoRI EP all2all backend, KV cache shuffle + paged attention for AITER"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/781

- config-keys:
    - kimik2.5-int4-b200-vllm
  description:
    - "Add Kimi-K2.5 INT4 vLLM benchmark for B200"
    - "Model: moonshotai/Kimi-K2.5 with --mm-encoder-tp-mode data and --trust-remote-code"
    - "Image: vllm/vllm-openai:v0.15.1"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/735

- config-keys:
    - gptoss-fp4-b200-vllm
    - gptoss-fp4-h100-vllm
    - gptoss-fp4-h200-vllm
  description:
    - "Update vLLM image from v0.13.0 to v0.15.1 for NVIDIA GPT-OSS configs"
    - "Remove deprecated async-scheduling flag (now enabled by default since v0.14.0)"
    - "Gains: CUTLASS MoE optimizations (~8% throughput), FP4 kernel improvements (~4% E2E on B200), torch.compile cold-start fix"
    - "v0.15.1 includes fix for prefix cache hit rate of 0% on GPT-OSS hybrid attention models"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/789
  
- config-keys:
    - dsr1-fp4-mi355x-atom
    - dsr1-fp4-mi355x-atom-mtp
  description:
    - "Update search-space configurations for DSR1 FP4 MI355X ATOM and ATOM-MTP"
    - "Comment out TP=4 configs, consolidate to TP=8 only"
    - "Extend concurrency range to conc-end: 256 across all sequence lengths (1k1k, 1k8k, 8k1k)"
    - "Fix MTP 1k8k conc-start from 256 to 4 to enable full concurrency sweep"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/699
 
- config-keys:
    - glm5-fp8-mi355x-sglang
  description:
    - "Add GLM-5 FP8 SGLang benchmark for MI355X"
    - "Model: zai-org/GLM-5-FP8 with NSA tilelang backends"
    - "Image: rocm/sgl-dev:v0.5.8.post1-rocm720-mi35x-20260219"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

- config-keys:
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
    - gptoss-fp4-mi355x-vllm
  description:
    - "Update AMD GPT-OSS vLLM images to v0.16.0 (MI300X/MI325X from v0.15.1, MI355X from custom v0.10.1)"
    - "MI355X: Fix env vars (VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION), add VLLM_ROCM_USE_AITER=1, remove deprecated flags"
    - "MI355X: Simplify compilation config to cudagraph_mode FULL_AND_PIECEWISE, add HIP_VISIBLE_DEVICES Ray fix"
    - "Gains: fused add+rmsnorm+pad for GPT-OSS (automatic via PassManager), AITER attention block size fix"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/806

- config-keys:
    - dsr1-fp4-mi355x-sglang
    - dsr1-fp8-mi300x-sglang
    - dsr1-fp8-mi325x-sglang
    - dsr1-fp8-mi355x-sglang
  description:
    - "Update SGLang image from v0.5.8 to v0.5.9 for AMD single-node DeepSeek R1 configs"
    - "Key changes: AITER v0.1.10.post3 with FP8 Prefill/Decode/KV Cache, FP8 prefill attention kernel, MORI EP two-batch overlapping, OOM fix for DeepSeek weight loading"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/816

- config-keys:
    - minimaxm2.5-fp8-h200-vllm
  description:
    - "Add MiniMax M2.5 FP8 single-node config for H200 with vLLM v0.16.0 (TP4)"
    - "New benchmark script with --trust-remote-code for MiniMaxAI/MiniMax-M2.5"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

- config-keys:
    - minimaxm2.5-fp8-mi325x-vllm
  description:
    - "Add MiniMax-M2.5 FP8 vLLM benchmark for MI325X"
    - "Model: MiniMaxAI/MiniMax-M2.5 with --trust-remote-code"
    - "Image: vllm/vllm-openai-rocm:v0.16.0"
    - "Environment: VLLM_ROCM_USE_AITER=1"
    - "TP=2 and TP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/836

- config-keys:
    - minimaxm2.5-fp8-mi300x-vllm
  description:
    - "Add MiniMax-M2.5 FP8 vLLM benchmark for MI300X"
    - "Model: MiniMaxAI/MiniMax-M2.5 with --trust-remote-code"
    - "Image: vllm/vllm-openai-rocm:v0.16.0"
    - "Environment: VLLM_ROCM_USE_AITER=1"
    - "TP=2 and TP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/837

- config-keys:
    - kimik2.5-fp4-mi355x-vllm
  description:
    - "Add Kimi-K2.5 MXFP4 vLLM benchmark for MI355X"
    - "Model: amd/Kimi-K2.5-MXFP4 with --mm-encoder-tp-mode data and --trust-remote-code"
    - "Image: vllm/vllm-openai-rocm:v0.15.1"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/825

- config-keys:
    - qwen3.5-bf16-mi325x-sglang
  description:
    - "Add Qwen3.5-397B-A17B BF16 SGLang benchmark for MI325X"
    - "Image: lmsysorg/sglang:v0.5.9-rocm720-mi30x"
    - "Uses triton attention backend, TP=8, concurrency 4-64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

- config-keys:
    - qwen3.5-bf16-mi300x-sglang
  description:
    - "Add Qwen3.5-397B-A17B BF16 SGLang benchmark for MI300X"
    - "Image: lmsysorg/sglang:v0.5.9-rocm720-mi30x"
    - "Uses triton attention backend, TP=8, concurrency 4-64"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/843

- config-keys:
    - qwen3.5-fp8-mi325x-sglang
  description:
    - "Add Qwen3.5-397B-A17B-FP8 SGLang benchmark for MI325X"
    - "Image: lmsysorg/sglang:v0.5.9-rocm720-mi30x"
    - "Following AMD Andy Luo's recipe with triton attention backend"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

- config-keys:
    - qwen3.5-fp8-mi300x-sglang
  description:
    - "Add Qwen3.5-397B-A17B-FP8 SGLang benchmark for MI300X"
    - "Image: lmsysorg/sglang:v0.5.9-rocm720-mi30x"
    - "Uses triton attention backend, TP=8, concurrency 4-64"
    - "Following AMD Andy Luo's recipe"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/850

- config-keys:
    - kimik2.5-int4-h200-vllm
  description:
    - "Add Kimi-K2.5 INT4 vLLM benchmark for H200"
    - "Model: moonshotai/Kimi-K2.5 with --reasoning-parser kimi_k2 and --trust-remote-code"
    - "Image: vllm/vllm-openai:v0.16.0"
    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
    - "following https://docs.vllm.ai/projects/recipes/en/latest/moonshotai/Kimi-K2.5.html"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/839
  
- config-keys:
    - dsr1-fp8-mi355x-sglang-disagg
    - dsr1-fp8-mi355x-sglang-disagg-mtp
    - dsr1-fp4-mi355x-sglang-disagg
    - dsr1-fp4-mi355x-sglang-disagg-mtp
  description:
    - "Add more sweep configs for MI355X FP8/FP4 Disagg"
    - "Add TP/DP/EP size < 8 support "
    - "Support DSR1-0528 MTP Disagg"
    - "Bump SGL mori image to Feb 27"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/823

- config-keys:
    - minimaxm2.5-fp8-h100-vllm
  description:
    - "Add MiniMax-M2.5 FP8 vLLM benchmark for H100"
    - "Model: MiniMaxAI/MiniMax-M2.5 with --trust-remote-code"
    - "Image: vllm/vllm-openai:v0.16.0"
    - "Switch from TP=8/EP=8 to TP=4/EP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k"
    - "Script uses conditional --enable-expert-parallel based on EP_SIZE env var"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/832

- config-keys:
    - qwen3.5-fp8-b200-sglang
  description:
    - "Add Qwen3.5-397B-A17B-FP8 SGLang benchmark configuration for B200"
    - "Image: lmsysorg/sglang:v0.5.9-cu129-amd64"
    - "Uses trtllm_mha attention backend and flashinfer_trtllm MOE runner"
    - "Enable SGLANG_ENABLE_FLASHINFER_GEMM=true, NCCL_NVLS_ENABLE=1"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/804

- config-keys: 
    - qwen3.5-fp8-h200-sglang
  description:
    - "Add Qwen 3.5 FP8 H200 SGLang configuration"
    - "Model: Qwen/Qwen3.5-397B-A17B-FP8, runner: h200, image: lmsysorg/sglang:v0.5.8-cu130-amd64"
    - "Benchmark script: benchmarks/single_node/qwen3.5_fp8_h200.sh"
    - "Server: reasoning-parser qwen3, tool-call-parser qwen3_coder, enable-flashinfer-allreduce-fusion, mem-fraction-static 0.8"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/855

- config-keys:
    - dsr1-fp8-mi355x-sglang
  description:
    - "Expanding TP search space"
    - "Adding kv-cache-fp8"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/865

- config-keys:
    - qwen3.5-bf16-b200-sglang
    - qwen3.5-bf16-mi300x-sglang
    - qwen3.5-bf16-mi325x-sglang
    - qwen3.5-bf16-mi355x-sglang
    - qwen3.5-fp8-b200-sglang
    - qwen3.5-fp8-h200-sglang
    - qwen3.5-fp8-mi300x-sglang
    - qwen3.5-fp8-mi325x-sglang
    - qwen3.5-fp8-mi355x-sglang
  description:
    - "Redo qwen eval"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/892
  evals-only: true
  
- config-keys:
    - gptoss-fp4-mi300x-vllm
    - gptoss-fp4-mi325x-vllm
    - gptoss-fp4-mi355x-vllm
  description:
    - "Update AMD GPT-OSS vLLM image from v0.16.0 to v0.17.0 for MI300X, MI325X, and MI355X"
    - "MI355X: Switch model to amd/gpt-oss-120b-w-mxfp4-a-fp8 (MXFP4 weights + FP8 activations)"
    - "MI355X: Add VLLM_ROCM_USE_AITER_TRITON_ROPE=1 for AITER triton RoPE kernel"
    - "Add AMDGCN_USE_BUFFER_OPS=0 and VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 env vars"
    - "Switch to --attention-backend ROCM_AITER_UNIFIED_ATTN and add fuse_rope_kvcache compilation pass"
    - "Remove deprecated VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION/VLLM_ROCM_USE_AITER_MHA env vars and compilation-config cudagraph_mode"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/867
  
- config-keys:
    - kimik2.5-fp4-b200-vllm
  description:
    - "Add Kimi K2.5 FP4 B200 vLLM benchmark configuration"
    - "Image: vllm/vllm-openai:v0.17.0"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/862

- config-keys:
    - minimaxm2.5-fp8-b200-vllm
  description:
    - "Add MiniMax-M2.5 FP8 vLLM benchmark for B200"
    - "Model: MiniMaxAI/MiniMax-M2.5 with --trust-remote-code"
    - "Image: vllm/vllm-openai:v0.17.0"
    - "TP=2 and TP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/757

- config-keys:
    - dsr1-fp4-mi355x-sglang-disagg
    - dsr1-fp4-mi355x-sglang-disagg-mtp
  description:
    - "Add more sweep configs for MI355X FP4 Disagg"
    - "Bump FP4 image to Feb 27 latest"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/899

- config-keys:
    - minimaxm2.5-fp8-h200-vllm
  description:
    - "Extend MiniMax M2.5 FP8 single-node config for H200 with vLLM v0.16.0 (TP8)"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/869
  
- config-keys:
    - dsr1-fp8-b200-dynamo-sglang
    - dsr1-fp8-b200-dynamo-sglang-mtp
  description:
    - "Update B200 FP8 DSR1 8k1k dynamo-sglang recipes to new pareto configs"
    - "Replace old per-file recipes with resolved variants from consolidated 8k1k.yaml"
    - "14 variants: STP/MTP x low-latency/max-throughput with updated concurrencies and scale points"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/907

- config-keys:
    - glm5-fp8-h200-sglang
  description:
    - "Add GLM-5 FP8 SGLang H200 single-node benchmark"
    - "Model: zai-org/GLM-5-FP8, image: lmsysorg/sglang:glm5-hopper"
    - "Benchmark script: benchmarks/single_node/glm5_fp8_h200.sh"
    - "Tool-call-parser glm47, reasoning-parser glm45, mem-fraction-static 0.85"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/914
  
- config-keys:
    - glm5-fp8-b200-sglang
  description:
    - "Add GLM-5 FP8 SGLang benchmark for B200"
    - "Supports TP8 (low latency) and DEP8 (high throughput) modes with NSA attention backend"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/915
  
- config-keys:
    - qwen3.5-fp8-b200-sglang-mtp
  description:
    - "Add Single Node Agg FP8 MTP config for Qwen3.5 B200 SGLang"
    - "EAGLE speculative decoding: num-steps 3, draft-tokens 4, topk 1"
    - "New script: benchmarks/single_node/qwen3.5_fp8_b200_mtp.sh"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/898

- config-keys:
    - minimaxm2.5-fp8-mi355x-vllm
  description:
    - "ADD minimax TP=8 with EP, in config of 1k1k, 1k8k, and 8k1k sequence lengths"
    - "Config concurrency: 32-256"
    - "update image to vllm/vllm-openai-rocm:v0.18.0"
  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/927