Files
sglang/scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/2p1d-ep16.yaml
T

96 lines
4.4 KiB
YAML

# MI355X DeepSeek-V4-Flash FP4 4-node 2P1D disaggregation recipe — narrow-prefill EP8
# + wide-decode EP16 (Oren's config: wide EP only helps decode).
#
# Two prefill engines (EP8, one node each; the router fans requests across both) +
# one decode engine (EP16) spanning 2 nodes. Still one logical P/D pair per role
# group, 4 nodes total. nodes-per-engine = ceil(TP/8): prefill 8->1, decode 16->2,
# so the launcher emits cross-node --nnodes/--node-rank/--dist-init-addr for the
# decode engine only. Prefill EP8 keeps MoE all-to-all INTRA-node (XGMI, known
# good); decode gets wide EP16 across nodes. Mismatched-TP KV (prefill TP8 ->
# decode TP16) is layout-compatible for DeepSeek MLA (KV is a replicated per-token
# latent, not head-sharded), and is carried over mori.
#
# Consumed by:
# * scripts/ci/slurm/process_result.py reads `resources` and
# `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
# * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
resources:
prefill_workers: 2
decode_workers: 1
backend:
sglang_config:
prefill:
tensor-parallel-size: 8
expert-parallel-size: 8
data-parallel-size: 8
decode:
tensor-parallel-size: 16
expert-parallel-size: 16
data-parallel-size: 16
runtime:
image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
attention_backend: dsv4
# RoCE HCAs (8/node) for mori MoE all-to-all AND the P->D KV transfer.
ib_devices: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
# Wide-EP MoE all-to-all backend (cross-node expert dispatch/combine).
moe_a2a_backend: mori
# KV P->D transfer backend (mori for both a2a and KV on this cluster).
kv_transfer_backend: mori
# Cross-node torch-distributed NIC for the wide decode engine's dist init.
dist_socket_ifname: eno0
# rocm720 0715 image needs the ROCm-7.0.0-alpha path OFF (validated).
rocm700a: 0
prefill_port: 30025
decode_port: 30026
prefill_bootstrap_port: 8998
decode_bootstrap_port: 9001
lb_port: 8000
# Base defaults; the wide_ep block overrides mem-fraction / max-req per role.
mem_fraction_static: 0.90
page_size: 256
max_running_requests: 256
chunked_prefill_size: 131072
swa_full_tokens_ratio: 0.1
# Per-role wide-EP tuning (validated 2P1D run, job 13196). Prefill and decode
# differ, so these split what the single-node EP<=8 recipes leave symmetric.
wide_ep:
kv_cache_dtype: fp8_e4m3
prefill_mem_fraction_static: 0.8
decode_mem_fraction_static: 0.85
prefill_chunked_prefill_size: 131072
prefill_max_running_requests: 1024
decode_max_running_requests: 1024
common_extra_flags: "--moe-dense-tp-size 1 --enable-dp-lm-head --decode-log-interval 100 --watchdog-timeout 3600 --load-balance-method round_robin"
prefill_extra_flags: "--context-length 9217 --max-total-tokens 262144"
# Cap the decode KV pool: the ionic NIC rejects any single RDMA memory
# region over 4 GiB and mori registers each KV buffer as one region
# (no chunking). Largest buffer = max_total_num_tokens * 256 B, so the
# ceiling is 16,777,216 tokens; 16,000,000 leaves headroom.
decode_extra_flags: "--max-total-tokens 16000000 --cuda-graph-bs-decode 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128"
prefill_extra_env:
MORI_MAX_DISPATCH_TOKENS_PREFILL: 8192
MORI_MAX_DISPATCH_TOKENS_DECODE: 256
SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 16384
decode_extra_env:
MORI_MAX_DISPATCH_TOKENS_DECODE: 64
MORI_MOE_MAX_INPUT_TOKENS_DECODE: 332
SGLANG_MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK: 128
bench:
# bench_serving --max-concurrency sweep; one result JSON per concurrency.
concurrencies: [1, 8, 16, 32, 64, 128, 256]
num_prompts_factor: 4 # num-prompts = concurrency * factor
random_range_ratio: 1.0
# Correctness gate run through the PD path before the perf sweep (full GSM8K,
# 8-shot, accuracy > 0.91). A regression here fails the nightly even when
# throughput looks fine ("fast but wrong").
accuracy:
enabled: true
num_shots: 8
num_questions: 1319 # full GSM8K test set
threshold: 0.91