123 lines
4.8 KiB
YAML
123 lines
4.8 KiB
YAML
# MI355X Kimi-K2.6 (MXFP4 experts, FP8 KV) 4-node 2P1D disaggregation recipe (MTP) — narrow-prefill EP8
|
|
# + wide-decode EP16 (mirrors the DSV4-Pro Oren 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); decode
|
|
# gets wide EP16 across nodes. KV (prefill TP8 -> decode TP16) is carried over mori.
|
|
#
|
|
# Kimi-specific bits vs the DSV4-Pro EP16 recipe: split attention backends
|
|
# (aiter prefill / triton decode), the Kimi model env + parsers, and the Kimi
|
|
# model path. Everything else (2P1D topology, mori a2a + KV, dist init) is shared.
|
|
#
|
|
# 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`, `model`, `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
|
|
|
|
# Model-specific docker env + sglang server args (written verbatim via
|
|
# model_flags.sh). Each server arg + value is a SEPARATE list item.
|
|
model:
|
|
env:
|
|
SGLANG_USE_AITER: 1
|
|
SGLANG_ROCM_FUSED_DECODE_MLA: 0
|
|
server_args:
|
|
- --model-loader-extra-config
|
|
- '{"enable_multithread_load": true}'
|
|
- --reasoning-parser
|
|
- kimi_k2
|
|
- --tool-call-parser
|
|
- kimi_k2
|
|
|
|
runtime:
|
|
image: lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260715
|
|
# Kimi uses split attention backends (aiter prefill / triton decode), not a
|
|
# single --attention-backend.
|
|
prefill_attention_backend: aiter
|
|
decode_attention_backend: triton
|
|
# 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 (starting point cloned from the validated DSV4-Pro
|
|
# EP16 run; may need Kimi-specific retuning). Prefill and decode differ.
|
|
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"
|
|
decode_extra_flags: "--disable-cuda-graph"
|
|
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
|
|
|
|
|
|
# EAGLE3 speculative decoding with the external Kimi-K2.6 draft model (same
|
|
# draft + hyperparams as the validated EP8 1p1d-mtp.yaml). Decode stays eager
|
|
# (--disable-cuda-graph); MXFP4 experts run the aiter MoE path (no int4 route-1
|
|
# dense-compaction), so cuda-graph could be revisited once the leg is green.
|
|
mtp:
|
|
enabled: true
|
|
algorithm: EAGLE3
|
|
num_steps: 3
|
|
eagle_topk: 1
|
|
num_draft_tokens: 4
|
|
draft_model_path: /it-share/model_coverage/models--lightseekorg--kimi-k2.6-eagle3.1-mla
|
|
|
|
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 > 0.92). Mirrors the single-node Kimi-K2.6 eval threshold.
|
|
accuracy:
|
|
enabled: true
|
|
num_shots: 8
|
|
num_questions: 1319 # full GSM8K test set
|
|
threshold: 0.92
|