disagg_same_gpu.sh 3.84 KB
Newer Older
1
#!/bin/bash
2
# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
3
4
# SPDX-License-Identifier: Apache-2.0
#
5
6
7
8
9
10
11
12
13
14
15
# Disaggregated prefill/decode on a SINGLE GPU.
# Per-worker VRAM is estimated from model parameters below. Override individual
# knobs (MAX_MODEL_LEN, MAX_CONCURRENT_SEQS) via env vars, or set
# DYN_GPU_MEMORY_FRACTION_OVERRIDE to bypass the calculation entirely.
#
# Measured reference (Qwen/Qwen3-0.6B, --max-model-len 4096, RTX 6000 Ada 48 GiB):
#   estimate (from gpu_utils.sh) : ~4.0 GiB per worker (~8.0 GiB total)
#   actual (nvidia-smi)          : ~3.4 GiB per worker (~6.7 GiB total)
#   fraction per worker (for 48 GiB) : 0.09
#   The ~1.3 GiB pad comes from the overhead term (CUDA ctx + activations).
#   Overestimating is intentional -- better to pad than OOM.
16

17
18
19
set -e
trap 'echo Cleaning up...; kill 0' EXIT

20
21
SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
source "$SCRIPT_DIR/../../../common/gpu_utils.sh"
22

23
MODEL="Qwen/Qwen3-0.6B"
24

25
26
27
# ---- Tunable (override via env vars) ----
MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
28

29
30
31
# ---- Estimate per-worker VRAM (see examples/common/gpu_utils.md) ----
# Sets _EW_WEIGHTS_GIB, _EW_KV_GIB, _EW_OVERHEAD_GIB, _EW_TOTAL_GIB
estimate_worker_vram "$MODEL" "$MAX_MODEL_LEN" "$MAX_CONCURRENT_SEQS" vllm
32

33
34
35
36
# DYN_GPU_MEMORY_FRACTION_OVERRIDE takes precedence (profiler binary search).
# In single-GPU mode, split the override evenly between the two workers.
if [[ -n "${DYN_GPU_MEMORY_FRACTION_OVERRIDE:-}" ]]; then
    GPU_MEM_FRACTION=$(awk -v f="$DYN_GPU_MEMORY_FRACTION_OVERRIDE" 'BEGIN { printf "%.2f", f / 2 }')
37
else
38
    GPU_MEM_FRACTION=$(gpu_worker_fraction vllm)
39
40
fi

41
source "$SCRIPT_DIR/../../../common/launch_utils.sh"
42

43
HTTP_PORT="${DYN_HTTP_PORT:-8000}"
44
45
46
47
print_launch_banner "Launching Disaggregated on Same GPU (1 GPU)" "$MODEL" "$HTTP_PORT" \
    "Max seq len: $MAX_MODEL_LEN" \
    "GPU Mem:     ${GPU_MEM_FRACTION} per worker (~${_EW_TOTAL_GIB} GiB each)" \
    "  estimate:  weights=${_EW_WEIGHTS_GIB} + kv=${_EW_KV_GIB} + overhead=${_EW_OVERHEAD_GIB} GiB"
48

49
# run ingress
50
51
# dynamo.frontend accepts either --http-port flag or DYN_HTTP_PORT env var (defaults to 8000)
python3 -m dynamo.frontend &
52
53
54

# run decode worker with metrics on port 8081
# --enforce-eager is added for quick deployment. for production use, need to remove this flag
55
56
57
# For disaggregated deployments we standardize on DYN_SYSTEM_PORT1/2 instead of
# *_PREFILL/*_DECODE env names so test harnesses can set one simple pair.
DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
58
59
CUDA_VISIBLE_DEVICES=0 \
python3 -m dynamo.vllm \
60
  --model "$MODEL" \
61
  --enforce-eager \
62
  --disaggregation-mode decode \
63
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
64
65
  --gpu-memory-utilization "${GPU_MEM_FRACTION}" \
  --max-model-len "$MAX_MODEL_LEN" &
66
67
68
69
70
71
72
73
74
75

# Wait for decode worker to initialize before starting prefill worker
# This prevents both workers from competing for GPU memory simultaneously, which can cause OOM.
# The decode worker needs time to:
# 1. Load model weights and allocate its memory fraction
# 2. Initialize KV cache
# 3. Register with NATS service discovery so prefill worker can find it
echo "Waiting for decode worker to initialize..."
sleep 10

76
# run prefill worker with metrics on port 8082
77
DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT2:-8082} \
78
79
80
VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \
CUDA_VISIBLE_DEVICES=0 \
python3 -m dynamo.vllm \
81
  --model "$MODEL" \
82
  --enforce-eager \
83
  --disaggregation-mode prefill \
84
  --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}' \
85
86
  --gpu-memory-utilization "${GPU_MEM_FRACTION}" \
  --max-model-len "$MAX_MODEL_LEN" \
87
88
89
90
  --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20081","enable_kv_cache_events":true}' &

# Exit on first worker failure; kill 0 in the EXIT trap tears down the rest
wait_any_exit