[NV] Add gb300 nvfp4 qwen3.5 MTP 8k1k recipes

benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen1tep16_mtp3_c48.yaml

@@ -0,0 +1,167 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/1ctxDEP2c128_1genTEP16c32_E2E_C32/conc32
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx1dep2_gen1tep16_mtp3_c48
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 40
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "48"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true

benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen4tp16_mtp3_c4.yaml

@@ -0,0 +1,156 @@
+# Derived from config #01 for one-request-per-decode-worker minimum-TPOT measurement.
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/1ctxDEP2c128_4genTP16c4_E2E_C16/conc16
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx1dep2_gen4tp16_mtp3_c4
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 16
+  decode_workers: 4
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 1
+      max_num_tokens: 4
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true