vLLM DSv4 GB200 PD 042326

configs/patches/vllm_numa_bind_hash_fix.py

@@ -0,0 +1,84 @@
+"""
+Patch vLLM's ParallelConfig.compute_hash to exclude NUMA-bind fields
+(numa_bind / numa_bind_nodes / numa_bind_cpus) from the DP consistency hash.
+
+Symptom (seen on GB300, 1 worker, DP=4, numa-bind=True):
+    RuntimeError: Configuration mismatch detected for engine 3.
+    All DP workers must have identical configurations for parameters that
+    affect collective communication ...
+
+Root cause: when numa-bind is enabled, each DP rank auto-detects and stores
+its own per-rank NUMA node in ParallelConfig.numa_bind_nodes. These per-rank
+values enter compute_hash(), so ranks on different NUMA nodes produce
+different hashes and fail the DP startup check. NUMA binding affects only
+host-side memory locality, not collective-communication semantics, so it is
+safe to exclude from the DP hash.
+
+Reference: vllm/config/parallel.py, ParallelConfig.compute_hash(),
+ignored_factors set.
+"""
+
+import sys
+from pathlib import Path
+
+TARGET = Path(
+    "/usr/local/lib/python3.12/dist-packages/vllm/config/parallel.py"
+)
+
+# Idempotency: if any of our additions is already present, skip.
+MARKER = '"numa_bind",'
+
+# Anchor: the last entry of the existing ignored_factors set in the
+# upstream compute_hash method. We insert the three numa fields just
+# before the closing brace.
+OLD = '            "_api_process_rank",\n        }'
+
+NEW = (
+    '            "_api_process_rank",\n'
+    '            # srt-slurm-sa hotfix: numa-bind fields are per-rank runtime\n'
+    '            # topology, not collective-communication semantics.\n'
+    '            "numa_bind",\n'
+    '            "numa_bind_nodes",\n'
+    '            "numa_bind_cpus",\n'
+    '        }'
+)
+
+
+def main():
+    if not TARGET.exists():
+        print(f"[vllm-numa-bind-hash-fix] Target not found: {TARGET}", file=sys.stderr)
+        sys.exit(1)
+
+    content = TARGET.read_text()
+
+    if MARKER in content:
+        print("[vllm-numa-bind-hash-fix] Already patched, skipping.", file=sys.stderr)
+        return
+
+    count = content.count(OLD)
+    if count == 0:
+        print(
+            "[vllm-numa-bind-hash-fix] Could not find ignored_factors anchor. "
+            "vLLM version may have drifted; inspect ParallelConfig.compute_hash().",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+    if count > 1:
+        print(
+            f"[vllm-numa-bind-hash-fix] Anchor is ambiguous ({count} occurrences); "
+            "refusing to patch.",
+            file=sys.stderr,
+        )
+        sys.exit(1)
+
+    content = content.replace(OLD, NEW)
+    TARGET.write_text(content)
+    print(
+        "[vllm-numa-bind-hash-fix] Added numa_bind/numa_bind_nodes/numa_bind_cpus "
+        "to ParallelConfig.compute_hash ignored_factors.",
+        file=sys.stderr,
+    )
+
+
+if __name__ == "__main__":
+    main()

configs/vllm-container-deps.sh

@@ -2,4 +2,8 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-pip install msgpack
+pip install msgpack
+
+if [ -f /configs/patches/vllm_numa_bind_hash_fix.py ]; then
+    python3 /configs/patches/vllm_numa_bind_hash_fix.py
+fi

recipes/vllm/deepseek-v4-pro/8k1k/disagg-gb200-1p1d-dep8-dep8-16-c256-c512-c1024-offload.yaml

@@ -0,0 +1,114 @@
+name: "svf-vllm-disagg-gb200-2p1d-dep8-dep16"
+model:
+  path: "deepseekv4-fp4"
+  container: "vllm/vllm-openai@sha256:2af012a17c2cee0bc1428c03a8a5e42b552f25dc6f73495ab5a29ccf4123c257"
+  precision: "fp4"
+
+dynamo:
+  version: 1.0.2
+  install: true
+
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_P2P_LEVEL: NVL
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_P2P_LEVEL: NVL
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enforce-eager: true
+      max-model-len: 16384
+      max-num-seqs: 16
+      max-num-batched-tokens: 32768
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.8
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: true
+      offload-group-size: 3
+      offload-num-in-group: 1
+      offload-prefetch-step: 2
+      # offload-params: "w13_weight w2_weight w13_weight_scale w2_weight_scale wq_b wo_a wo_b shared_experts"
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      max-model-len: 16384
+      max-num-seqs: 128
+      max-cudagraph-capture-size: 128
+      max-num-batched-tokens: 128
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4x8x16x32x64x256x512x1024"
+  req_rate: "inf"

recipes/vllm/deepseek-v4-pro/8k1k/disagg-gb200-3p1d-dep8-dep16-40-c4096-offload.yaml

@@ -0,0 +1,113 @@
+name: "svf-vllm-disagg-gb200-2p1d-dep8-dep16"
+model:
+  path: "deepseekv4-fp4"
+  container: "vllm/vllm-openai@sha256:2af012a17c2cee0bc1428c03a8a5e42b552f25dc6f73495ab5a29ccf4123c257"
+  precision: "fp4"
+
+dynamo:
+  version: 1.0.2
+  install: true
+
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 6
+  decode_nodes: 4
+  prefill_workers: 3
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 16
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_P2P_LEVEL: NVL
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_TLS: "cuda_copy,cuda_ipc,tcp"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_P2P_LEVEL: NVL
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enforce-eager: true
+      max-model-len: 16384
+      max-num-seqs: 16
+      max-num-batched-tokens: 32768
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.8
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: true
+      offload-group-size: 3
+      offload-num-in-group: 1
+      offload-prefetch-step: 2
+      # offload-params: "w13_weight w2_weight w13_weight_scale w2_weight_scale wq_b wo_a wo_b shared_experts"
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 16
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      max-model-len: 16384
+      max-num-seqs: 256
+      max-cudagraph-capture-size: 256
+      max-num-batched-tokens: 256
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "3072x4096"
+  req_rate: "inf"