[Klaud Cold][agentic experiment][Variant F] Kimi-K3 B200 agg TP8xPP2 agentic — SimpleCPU offload + prefix-cache retention / Kimi-K3 B200 聚合式 TP8xPP2 智能体实验——SimpleCPU 卸载 + 前缀缓存留存

MODELS.md

@@ -22,7 +22,7 @@ This document tracks every model benchmarked by InferenceX-e2e: when it was adde
 | Model architecture class | Prefix | Date added | Active scenarios | Deprecated scenarios |
 |---|---|---|---|---|
 | Qwen3.8 2.4T | `qwen3.8` | TBD | Agentic coding | |
-| Kimi-K3 | `kimik3` | 2026-07-27 | Agentic coding | |
+| Kimi-K3 | `kimik3` | 2026-07-27 ([#2372](https://github.com/SemiAnalysisAI/InferenceX/pull/2372)) | Agentic coding | |
 | GLM-5.2 | `glm5.2` | 2026-07-18 ([#2268](https://github.com/SemiAnalysisAI/InferenceX/pull/2268)) | Agentic coding | |
 | MiniMax-M3 | `minimaxm3` | 2026-06-12 ([#1724](https://github.com/SemiAnalysisAI/InferenceX/pull/1724)) | Single-turn 8k1k, Agentic coding | Single-turn 1k1k |
 | DeepSeek-V4-Pro | `dsv4` | 2026-04-24 ([#1130](https://github.com/SemiAnalysisAI/InferenceX/pull/1130)) | Single-turn 8k1k, Agentic coding | Single-turn 1k1k |

MODELS_zh.md

@@ -22,7 +22,7 @@
 | 模型架构类别 | 前缀 | 加入日期 | 启用场景 | 已弃用场景 |
 |---|---|---|---|---|
 | Qwen3.8 2.4T | `qwen3.8` | 待定 | 智能体编码 | |
-| Kimi-K3 | `kimik3` | 2026-07-27 | 智能体编码 | |
+| Kimi-K3 | `kimik3` | 2026-07-27 ([#2372](https://github.com/SemiAnalysisAI/InferenceX/pull/2372)) | 智能体编码 | |
 | GLM-5.2 | `glm5.2` | 2026-07-18（[#2268](https://github.com/SemiAnalysisAI/InferenceX/pull/2268)） | 智能体编码 | |
 | MiniMax-M3 | `minimaxm3` | 2026-06-12（[#1724](https://github.com/SemiAnalysisAI/InferenceX/pull/1724)） | 单轮 8k1k、智能体编码 | 单轮 1k1k |
 | DeepSeek-V4-Pro | `dsv4` | 2026-04-24（[#1130](https://github.com/SemiAnalysisAI/InferenceX/pull/1130)） | 单轮 8k1k、智能体编码 | 单轮 1k1k |

benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-container-deps.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# Setup script for the Kimi-K3 vLLM bring-up image (vllm/vllm-openai:kimi-k3).
+# srt-slurm runs this in every worker container before dynamo install and
+# worker startup (recipe field: setup_script).
+
+set -euo pipefail
+
+# The image's first decode step crashes in the KDA hybrid-state postprocess:
+#   vllm/v1/worker/gpu/model_states/mamba_hybrid.py, postprocess_state:
+#   IndexError: index_fill_(): Expected dtype int64 for index.
+# torch's index_fill_ requires an int64 index tensor, but the runner passes
+# the int32 idx_mapping (hit by moonshotai/Kimi-K3 agentic bring-up, first
+# decode step, engine v0.1.dev19262+gb6bbf29dd). Coerce the index to int64.
+# Idempotent: exits 0 if the patch is already applied.
+python3 - <<'PY'
+import pathlib
+import re
+
+import vllm.v1.worker.gpu.model_states.mamba_hybrid as mh
+
+path = pathlib.Path(mh.__file__)
+src = path.read_text()
+if "idx_mapping.long()" in src:
+    print(f"mamba_hybrid index_fill_ patch already applied: {path}")
+    raise SystemExit(0)
+
+new, n = re.subn(
+    r"index_fill_\(\s*0,\s*idx_mapping,",
+    "index_fill_(0, idx_mapping.long(),",
+    src,
+)
+if n != 1:
+    raise SystemExit(
+        f"expected exactly one index_fill_(0, idx_mapping, ...) call in "
+        f"{path}, found {n} — image layout changed, refusing to patch"
+    )
+path.write_text(new)
+print(f"Patched mamba_hybrid index_fill_ index dtype: {path}")
+PY

benchmarks/multi_node/srt-slurm-recipes/patches/srt-slurm-pr278-direct-vllm-multinode.patch

@@ -0,0 +1,62 @@
+diff --git a/src/srtctl/backends/vllm.py b/src/srtctl/backends/vllm.py
+index 74f673b..377606a 100644
+--- a/src/srtctl/backends/vllm.py
++++ b/src/srtctl/backends/vllm.py
+@@ -716,25 +716,31 @@ class VLLMProtocol:
+         if frontend_type == "vllm":
+             if mode != "agg":
+                 raise ValueError("frontend.type: vllm supports aggregate vLLM jobs only")
+-            if is_multi_node:
+-                raise ValueError("frontend.type: vllm currently supports single-node aggregate jobs only")
+
+             config.pop("host", None)
+             config.pop("port", None)
+             config.pop("connector", None)
+             config.setdefault("served-model-name", served_model_name)
+
+-            cmd.extend(
+-                [
+-                    "vllm",
+-                    "serve",
+-                    model_arg,
+-                    "--host",
+-                    "0.0.0.0",
+-                    "--port",
+-                    str(runtime.frontend_port),
+-                ]
+-            )
++            node_rank = endpoint_nodes.index(process.node)
++            cmd.extend(["vllm", "serve", model_arg])
++            if node_rank == 0:
++                cmd.extend(["--host", "0.0.0.0", "--port", str(runtime.frontend_port)])
++            if is_multi_node:
++                # vLLM-native multi-node serve (torchrun-style): the leader owns
++                # the OpenAI server; other node ranks run headless engine workers.
++                cmd.extend(
++                    [
++                        "--master-addr",
++                        leader_ip,
++                        "--nnodes",
++                        str(len(endpoint_nodes)),
++                        "--node-rank",
++                        str(node_rank),
++                    ]
++                )
++                if node_rank > 0:
++                    cmd.append("--headless")
+             if not self.set_cuda_visible_devices:
+                 device_ids = ",".join(str(i) for i in sorted(process.gpu_indices))
+                 if device_ids:
+diff --git a/src/srtctl/core/schema.py b/src/srtctl/core/schema.py
+index 1263ddc..0ef7ae4 100644
+--- a/src/srtctl/core/schema.py
++++ b/src/srtctl/core/schema.py
+@@ -1587,8 +1587,6 @@ class SrtConfig:
+             raise ValidationError("frontend.type: vllm supports aggregate jobs only, not disaggregated layouts")
+         if self.resources.num_agg < 1:
+             raise ValidationError("frontend.type: vllm requires resources.agg_workers >= 1")
+-        if (self.resources.agg_nodes or 1) != 1:
+-            raise ValidationError("frontend.type: vllm currently supports single-node aggregate jobs only")
+
+     def _validate_het_jobs(self):
+         """When ``resources.het_jobs`` is set to True, enforce supported shape.

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-agentic.yaml

@@ -0,0 +1,183 @@
+name: "kimik3-vllm-agg-b200-tp8pp2-simplecpu-retention-agentic"
+
+# Kimi-K3 MXFP4 B200 AGGREGATED TP8 x PP2 agentic recipe (2 nodes / 16 GPUs).
+# The native MXFP4 checkpoint (2.8T total params, ~1.4TB of weights) does not
+# fit one 8xB200 node, so TP8 shards attention/dense (/8) and PP2 splits the
+# 93 layers (/2) across 16 GPUs. Plain TP (NOT TEP): expert parallelism is
+# deliberately off, so the 896 routed experts are TP-sharded inside each
+# pipeline stage. Node allocation = tp*pp/gpus_per_node = 8*2/8 = 2 nodes.
+# Aggregated (single worker, decode num-worker 0) — no P/D split, no NIXL.
+# VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION fuses the K3 LatentMoE tail path in
+# the kimi-k3 bring-up image.
+model:
+  path: "kimik3"
+  container: "vllm/vllm-openai:kimi-k3"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:kimi-k3"
+
+# Direct vLLM serving (frontend.type: vllm, srt-slurm PR #278 + the
+# InferenceX multinode patch): `vllm serve` owns the OpenAI port itself, so
+# no Dynamo frontend/worker is involved and no dynamo install is needed.
+dynamo:
+  install: false
+
+# Patches the image's mamba_hybrid postprocess_state: torch index_fill_
+# requires an int64 index but the runner passes the int32 idx_mapping,
+# crashing the first decode step (IndexError: Expected dtype int64 for index).
+setup_script: kimi-k3-container-deps.sh
+
+environment:
+  # SimpleCPUOffloadConnector: identical prefixes must hash to identical
+  # block keys across ranks.
+  PYTHONHASHSEED: "42"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 1440
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  # Direct vLLM OpenAI server (srt-slurm PR #278): the vllm serve leader owns
+  # the public port; rank-1 runs a headless engine worker (vLLM-native
+  # multi-node TP8xPP2 via --master-addr/--nnodes/--node-rank, enabled by
+  # patches/srt-slurm-pr278-direct-vllm-multinode.patch).
+  type: vllm
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    # ~1.4TB of MXFP4 weights off shared Lustre: keep the engine-ready window
+    # generous, and let one long AgentX request hold a PP stage beyond vLLM's
+    # 300-second model-execution default.
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    # Allocator-fragmentation fix proven on the identical SimpleCPU connector
+    # config in #2370 (variant E, closed): a 2.36 GiB MLA prefill transient
+    # OOM'd while 2.63 GiB sat reserved-but-unallocated. The GPU-resident
+    # variant D runs green without this; the offload connector's staging
+    # shifts the layout enough to fragment. NCCL_CUMEM_ENABLE unchanged.
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    # Prefix-cache retention (variant F): 0. Any positive value must be a
+    # multiple of Kimi-K3's KDA-hybrid scheduler_block_size (3145728; the GB
+    # recipes' 32768 is hard-rejected at engine init — verified on this PR
+    # family), and 0 is the only value below one 3.1M-token scheduler block.
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    # No VLLM_PREFIX_CACHE_RETENTION_INTERVAL: the GB200/GB300 AgentX value
+    # (32768) hard-fails engine init on Kimi-K3 — the KDA hybrid gives it a
+    # scheduler_block_size of 3145728 and the interval must be a multiple of
+    # it ("VLLM_PREFIX_CACHE_RETENTION_INTERVAL (32768) must be non-negative
+    # and a multiple of scheduler_block_size (3145728)"). Default retention
+    # served fine in earlier runs.
+    NCCL_CUMEM_ENABLE: "1"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 2
+      # Required by vLLM's config validation when SimpleCPUOffloadConnector
+      # is combined with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True:
+      # torch's VMM allocator can remap KV virtual addresses under pinned/
+      # registered connector memory, so KV allocations must route through
+      # CuMemAllocator's pool (where expandable_segments is auto-disabled for
+      # KV while remaining active for transient activations). Same setting as
+      # the GB300 aggregate AgentX recipe.
+      enable-cumem-allocator: true
+      # Paired with the cumem allocator, as in the GB300 aggregate AgentX
+      # recipe: vLLM's custom all-reduce registers IPC buffers that the
+      # VMM-backed CuMemAllocator invalidates — CUDA-graph capture dies with
+      # "custom_all_reduce.cuh 'invalid argument'" (ninth sweep attempt).
+      # TP all-reduce falls back to NCCL.
+      disable-custom-all-reduce: true
+      trust-remote-code: true
+      # SimpleCPUOffloadConnector (variant E): offload KV blocks to CPU DRAM.
+      # enable_cross_layers_blocks false: with true, KV-cache init crashes on
+      # K3's KDA hybrid — "shape '[64980, 64, 576]' is invalid for input of
+      # size 199618560", a 12x mismatch equal to the MLA layers per PP stage
+      # (24/2); the cross-layer block folding gets the hybrid geometry wrong.
+      # The GB200 recipes also run connectors with cross-layer blocks off.
+      # cpu_bytes_to_use_per_rank = TOTAL_CPU_DRAM_GB * 1e9 / GPU_COUNT with
+      # the framework's own budget math for cluster:b200-dgxc — DRAM
+      # min(3,095,781 MiB, 2,861,022 MiB cap) * 0.80 utilization = 2399 GB
+      # per node (matrix total-cpu-dram-gb), / 8 ranks = 299,875,000,000
+      # bytes per rank.
+      kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":299875000000,"lazy_offload":false,"enable_cross_layers_blocks":false}}'
+      # Explicit (default-on) — offloaded blocks are only useful if prefix
+      # caching reuses them across agentic turns.
+      enable-prefix-caching: true
+      load-format: fastsafetensors
+      moe-backend: auto
+      # 0.90, not 0.95: the flashinfer trtllm MXFP4 MoE kernel allocates a
+      # ~1.6 GiB runtime workspace OUTSIDE vLLM's memory pool on the first
+      # forward; at 0.95 a 178 GiB B200 has only ~1.35 GiB free and the first
+      # warmup request OOMs (seen on the dynamo-frontend variants). 0.90
+      # matches the GB200/GB300 agentic recipes.
+      gpu-memory-utilization: 0.90
+      no-enable-flashinfer-autotune: true
+      # kimi_k3 parsers via the native vllm serve OpenAI-frontend flags —
+      # legitimate here because this recipe serves directly with vllm serve
+      # (frontend.type: vllm), not through the dynamo worker entrypoint that
+      # rejects them.
+      enable-auto-tool-choice: true
+      tool-call-parser: kimi_k3
+      reasoning-parser: kimi_k3
+      # No explicit max-model-len: let vLLM derive the native 1M window from
+      # the model config (agentic trajectories blow past any small cap, and
+      # K3's KDA layers keep per-token KV small — only the 24 gated-MLA
+      # layers hold cache). Prefix caching stays on (default) for trajectory
+      # reuse. Cap prefill chunks so a single long request cannot OOM a
+      # pipeline stage; let vLLM pick max-num-seqs.
+      max-num-batched-tokens: 8192
+
+sbatch_directives:
+  segment: "1"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    # Keep the aggregate worker in the multinode result schema so ingestion
+    # uses the zero decode-worker count instead of duplicating TP into P and D.
+    IS_MULTINODE: "true"
+    # aiperf's conv-aware routing emits nvext.session_control, a removed POC
+    # field this dynamo build 400-rejects at warmup (schema moved to
+    # router/routing_constraints/agent_hints). Same opt-out as the GB300
+    # aggregate AgentX recipes — and with a single aggregate worker there is
+    # no P/D routing to bind anyway.
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"

configs/nvidia-master.yaml

@@ -8357,3 +8357,54 @@ qwen3.5-fp8-gb200-dynamo-sglang-mtp:
           tp: 16
           ep: 16
           dp-attn: true
+
+# Kimi-K3 MXFP4 B200 aggregated vLLM via Dynamo (TP8 x PP2, 2 nodes / 16
+# GPUs), agentic bring-up. The native MXFP4 checkpoint (2.8T total params,
+# ~1.4TB weights) does not fit one 8xB200 node, so TP8 shards attention/dense
+# and PP2 splits layers. Plain TP (NOT TEP): ep 1, no expert parallelism —
+# the 896 routed experts are TP-sharded within each pipeline stage. Node
+# count = tp*pp/gpus_per_node = 8*2/8 = 2. Aggregated (prefill num-worker 1 +
+# decode num-worker 0, RECIPES.md section 5) — the single worker serves both
+# phases, so no P/D KV transfer. Dedicated kimi-k3 vLLM bring-up image with
+# VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1 and the kimi_k3 tool-call/reasoning
+# parsers.
+# Recipe: benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-agentic.yaml
+kimik3-fp4-b200-dynamo-vllm-agentic:
+  image: vllm/vllm-openai:kimi-k3
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:b200-dgxc
+  precision: fp4
+  # framework stays dynamo-vllm for launcher routing, but this variant serves
+  # DIRECTLY with vllm serve (srt-slurm PR #278 frontend.type: vllm + the
+  # InferenceX multinode patch) — no dynamo frontend/worker/router involved.
+  framework: dynamo-vllm
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    # Variant E: SimpleCPUOffloadConnector KV offload to CPU DRAM.
+    # dram-utilization 0.80 of the capped node DRAM = 2400 GB/node budget;
+    # the recipe hardcodes the resulting 300 GB/rank connector config.
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: none
+        kv-offloading: dram
+        kv-offload-backend: { name: vllm-simple, version: "kimi-k3" }
+        conc-list: [1, 2, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 8
+          pp: 2
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-agentic.yaml"
+        # The aggregate worker also performs decode; keep the decode worker
+        # count at zero so result aggregation counts the 16 GPUs only once.
+        decode:
+          num-worker: 0
+          tp: 8
+          pp: 2
+          ep: 1
+          dp-attn: false