feat(catalog): add GLM-5.3 Flash dual-Spark deployment

catalog/engines/vllm-glm53-dual-gb10.yaml

@@ -0,0 +1,162 @@
+kind: engine_asset
+_profile: vllm
+metadata:
+  name: vllm-glm53-dual-gb10
+  type: vllm
+  version: "0.1.dev20051-sm121-v8-cache-safety"
+  supported_formats: [safetensors]
+  supported_model_types: [llm, vlm]
+hardware:
+  gpu_arch: Blackwell
+  vram_min_mib: 4096
+image:
+  # Local distribution is intentional: the image contains the GLM-5.3 SM121,
+  # FP8-KV, Kpool prefix-cache, and block-table safety patch stack. Install the
+  # same image on both Fleet devices before applying the two-node scenario.
+  name: aima/glm53-flash
+  tag: sm121-v8-local
+  size_approx_mb: 28200
+  distribution: local
+  platforms: [linux/arm64]
+patterns:
+  - "^aima/glm53-flash:sm121-v8-local$"
+  - "glm53-flash-arm64-cu130"
+container:
+  devices:
+    - /dev/infiniband
+  network_mode: host
+  shm_size: 64gb
+  ulimits:
+    memlock: "-1:-1"
+    stack: "67108864:67108864"
+startup:
+  env:
+    HF_HUB_OFFLINE: "1"
+    TRANSFORMERS_OFFLINE: "1"
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    NCCL_NET: IB
+    NCCL_IB_DISABLE: "0"
+    NCCL_IB_ADDR_FAMILY: AF_INET
+    NCCL_IB_ROCE_VERSION_NUM: "2"
+    NCCL_IB_MERGE_NICS: "1"
+    NCCL_CROSS_NIC: "0"
+    NCCL_CUMEM_ENABLE: "0"
+    NCCL_IGNORE_CPU_AFFINITY: "1"
+    NCCL_NVLS_ENABLE: "0"
+    NCCL_DEBUG: WARN
+  accepted_config_keys:
+    - host
+    - port
+    - served_model_name
+    - trust_remote_code
+    - dtype
+    - tensor_parallel_size
+    - pipeline_parallel_size
+    - distributed_executor_backend
+    - nnodes
+    - node_rank
+    - master_addr
+    - master_port
+    - headless
+    - max_model_len
+    - max_num_seqs
+    - max_num_batched_tokens
+    - gpu_memory_utilization
+    - kv_cache_dtype
+    - kv_cache_memory
+    - block_size
+    - prefix_match_unit
+    - moe_backend
+    - enforce_eager
+    - enable_prefix_caching
+    - enable_prompt_tokens_details
+    - enable_chunked_prefill
+    - enable_auto_tool_choice
+    - tool_call_parser
+    - reasoning_parser
+    - chat_template
+    - default_chat_template_kwargs
+    - limit_mm_per_prompt
+    - generation_config
+    - speculative_config
+  default_args:
+    host: 0.0.0.0
+    port: 8000
+    served_model_name: GLM-5.3-Flash-NVFP4
+    trust_remote_code: true
+    dtype: auto
+    tensor_parallel_size: 2
+    pipeline_parallel_size: 1
+    distributed_executor_backend: mp
+    nnodes: 2
+    master_port: 29530
+    max_model_len: 262144
+    max_num_seqs: 2
+    max_num_batched_tokens: 8192
+    gpu_memory_utilization: 0.85
+    kv_cache_dtype: fp8_e4m3
+    kv_cache_memory: 4445787956
+    block_size: 2304
+    prefix_match_unit: 128
+    moe_backend: marlin
+    enforce_eager: true
+    enable_prefix_caching: true
+    enable_prompt_tokens_details: true
+    enable_chunked_prefill: true
+    enable_auto_tool_choice: true
+    tool_call_parser: glm47
+    reasoning_parser: glm45
+    chat_template: /models/chat_template.jinja
+    default_chat_template_kwargs:
+      enable_thinking: true
+    limit_mm_per_prompt:
+      image: 4
+      video: 1
+    generation_config: vllm
+    speculative_config:
+      method: mtp
+      num_speculative_tokens: 4
+  health_check:
+    path: /v1/models
+    timeout_s: 3600
+  warmup:
+    enabled: true
+    prompt: "Return the word ready."
+    max_tokens: 8
+    timeout_s: 180
+  # A cold load takes roughly 15 minutes. Generic recovery must not replace a
+  # healthy loading process after live unified-memory fit changes.
+  recovery:
+    enabled: false
+api:
+  protocol: openai
+  base_path: /v1
+  upstream_model: GLM-5.3-Flash-NVFP4
+amplifier:
+  features:
+    - tensor_parallel_2
+    - nvfp4
+    - fp8_kv_cache
+    - mtp_4
+    - prefix_caching
+    - fine_grained_prefix_matching_128
+    - prompt_cache_usage_telemetry
+    - chunked_prefill
+    - multimodal
+    - tool_calling
+    - roce
+  performance_gain: "Two-node GB10 TP=2 with FP8 KV, MTP-4, and 128-token prefix matching for GLM-5.3 Flash NVFP4"
+  resource_expansion:
+    cpu_offload: false
+    ssd_offload: false
+    npu_offload: false
+  performance_multiplier: 1.52
+  extends_resource_boundary: true
+  effective_vram_multiplier: 2.0
+time_constraints:
+  cold_start_s: [600, 1200]
+  model_switch_s: [600, 1200]
+runtime:
+  default: container
+  platform_recommendations:
+    linux/arm64: container

catalog/models/glm-5.3-flash-nvfp4.yaml

@@ -0,0 +1,76 @@
+kind: model_asset
+metadata:
+  name: glm-5.3-flash-nvfp4
+  type: vlm
+  family: glm
+  parameter_count: "320B"
+  released_at: "2026-08-26"
+  aliases:
+    - GLM-5.3-Flash-NVFP4
+    - LibertAIDAI/GLM-5.3-Flash-NVFP4
+ui:
+  role: deployable
+  display_note: "Two DGX Spark coding profile: NVFP4 weights, FP8 KV, MTP-4, tools, multimodal input, and 256K context"
+  display_note_zh: "双机 DGX Spark Coding 配置：NVFP4 权重、FP8 KV、MTP-4、工具调用、多模态与 256K 上下文"
+capabilities:
+  deployment_scenario: glm-5.3-flash-nvfp4-2node
+openclaw:
+  chat_provider: true
+storage:
+  formats: [safetensors]
+  default_path_pattern: "{{.DataDir}}/models/{{.Name}}"
+  sources:
+    - type: huggingface
+      repo: LibertAIDAI/GLM-5.3-Flash-NVFP4
+      format: safetensors
+      quantization: nvfp4
+    - type: local_path
+      path: ""
+variants:
+  - name: glm-5.3-flash-nvfp4-2node-gb10
+    hardware:
+      gpu_arch: Blackwell
+      vram_min_mib: 0
+      ram_min_mib: 98304
+      # One GB10 is present on each Fleet device; the scenario supplies two nodes.
+      gpu_count_min: 1
+      unified_memory: true
+    engine: vllm-glm53-dual-gb10
+    format: safetensors
+    source:
+      type: huggingface
+      repo: LibertAIDAI/GLM-5.3-Flash-NVFP4
+      format: safetensors
+      quantization: nvfp4
+    default_config:
+      max_model_len: 262144
+      max_num_seqs: 2
+      max_num_batched_tokens: 8192
+      gpu_memory_utilization: 0.85
+      served_model_name: GLM-5.3-Flash-NVFP4
+      tensor_parallel_size: 2
+      pipeline_parallel_size: 1
+      kv_cache_dtype: fp8_e4m3
+      kv_cache_memory: 4445787956
+      block_size: 2304
+      prefix_match_unit: 128
+      speculative_config:
+        method: mtp
+        num_speculative_tokens: 4
+      enable_prefix_caching: true
+      enable_prompt_tokens_details: true
+      enable_chunked_prefill: true
+      enable_auto_tool_choice: true
+      tool_call_parser: glm47
+      reasoning_parser: glm45
+      default_chat_template_kwargs:
+        enable_thinking: true
+      limit_mm_per_prompt:
+        image: 4
+        video: 1
+    expected_performance:
+      startup_time_s: 900
+      cold_start_time_s: 1200
+      vram_mib: 110000
+      max_context_tokens: 262144
+      notes: "Validated on two 128GB unified-memory GB10 nodes over direct RoCE. Requires the matching cache-safety image on both Fleet devices."

catalog/scenarios/glm-5.3-flash-nvfp4-2node.yaml

@@ -0,0 +1,112 @@
+kind: deployment_scenario
+metadata:
+  name: glm-5.3-flash-nvfp4-2node
+  description: "GLM-5.3 Flash NVFP4 across two DGX Spark GB10 devices using AIMA Fleet"
+target:
+  hardware_profile: nvidia-gb10-arm64
+
+inputs:
+  - name: worker_device
+    label: Worker Fleet device
+    description: "AIMA Fleet device ID for rank 1"
+    kind: device
+    required: true
+  - name: master_addr
+    label: Master fabric address
+    description: "Head-node IPv4 address reachable over the direct RoCE link"
+    kind: string
+    required: true
+  - name: head_host_ip
+    label: Head vLLM host IP
+    description: "Head-node address advertised by vLLM, normally the master fabric address"
+    kind: string
+    required: true
+  - name: worker_host_ip
+    label: Worker vLLM host IP
+    description: "Worker-node address on the direct RoCE link"
+    kind: string
+    required: true
+  - name: head_socket_ifname
+    label: Head fabric interface
+    description: "Head NCCL/Gloo interface name"
+    kind: string
+    required: true
+  - name: worker_socket_ifname
+    label: Worker fabric interface
+    description: "Worker NCCL/Gloo interface name"
+    kind: string
+    required: true
+  - name: head_ib_hca
+    label: Head IB HCA
+    description: "Head NCCL_IB_HCA value"
+    kind: string
+    required: true
+  - name: worker_ib_hca
+    label: Worker IB HCA
+    description: "Worker NCCL_IB_HCA value"
+    kind: string
+    required: true
+  - name: head_gid_index
+    label: Head RoCE GID index
+    description: "RoCEv2 GID index matching the head fabric address"
+    kind: string
+    required: true
+  - name: worker_gid_index
+    label: Worker RoCE GID index
+    description: "RoCEv2 GID index matching the worker fabric address"
+    kind: string
+    required: true
+
+deployments:
+  - id: worker
+    device: "{{.worker_device}}"
+    model: glm-5.3-flash-nvfp4
+    engine: vllm-glm53-dual-gb10
+    no_pull: true
+    role: worker
+    config:
+      node_rank: 1
+      headless: true
+      master_addr: "{{.master_addr}}"
+    env:
+      VLLM_HOST_IP: "{{.worker_host_ip}}"
+      NCCL_SOCKET_IFNAME: "{{.worker_socket_ifname}}"
+      TP_SOCKET_IFNAME: "{{.worker_socket_ifname}}"
+      GLOO_SOCKET_IFNAME: "{{.worker_socket_ifname}}"
+      NCCL_IB_HCA: "{{.worker_ib_hca}}"
+      NCCL_IB_GID_INDEX: "{{.worker_gid_index}}"
+    notes: "Rank 1 starts first and waits for the head rendezvous."
+
+  - id: head
+    device: local
+    model: glm-5.3-flash-nvfp4
+    engine: vllm-glm53-dual-gb10
+    no_pull: true
+    role: head
+    config:
+      node_rank: 0
+      master_addr: "{{.master_addr}}"
+    env:
+      VLLM_HOST_IP: "{{.head_host_ip}}"
+      NCCL_SOCKET_IFNAME: "{{.head_socket_ifname}}"
+      TP_SOCKET_IFNAME: "{{.head_socket_ifname}}"
+      GLOO_SOCKET_IFNAME: "{{.head_socket_ifname}}"
+      NCCL_IB_HCA: "{{.head_ib_hca}}"
+      NCCL_IB_GID_INDEX: "{{.head_gid_index}}"
+    notes: "Rank 0 exposes the OpenAI-compatible API on port 8000."
+
+startup_order:
+  - step: 1
+    deployment: worker
+    wait_for: ""
+    timeout_s: 0
+    notes: "Launch the headless worker before rank 0."
+  - step: 2
+    deployment: head
+    wait_for: health_check
+    timeout_s: 3600
+    notes: "Wait for model load, MTP initialization, and the API health check."
+
+memory_budget:
+  total_unified_mib: 262144
+  notes: "Two 128GB unified-memory nodes; the validated profile pins a 4.14 GiB FP8 KV cache on each rank."