[AMD][MI355X] DSv4 SGLang agentic: v0.5.14 image, router, HiCache#2146
[AMD][MI355X] DSv4 SGLang agentic: v0.5.14 image, router, HiCache#2146seungrokj wants to merge 8 commits into
Conversation
…th router and HiCache support Align the agentic launcher with the fixed-seq-len sibling: dsv4 attention backend, fp8_e4m3 KV cache, disable-radix-cache, DP-attention env exports (SGLANG_SHARED_EXPERT_TP1, DP_USE_GATHERV, etc.), and two-batch-overlap. Add SGLang router for DP-attention configs and HiCache KV offloading. Sweep conc=48 across TP8 +/- DPA +/- HiCache. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
seungrokj
requested review from
billishyahao,
chunfangamd and
yctseng0211
as code owners
|
Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase For PR verification, add the PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs 感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 如需进行 PR 验证，请为此 PR 添加 PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档 |
seungrokj
added
AMD
full-sweep-fail-fast
agentx
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| --swa-full-tokens-ratio 0.15 | ||
| --page-size 256 | ||
| --kv-cache-dtype fp8_e4m3 | ||
| --context-length "$MAX_MODEL_LEN" |
There was a problem hiding this comment.
🔴 The launcher references "$MAX_MODEL_LEN" in --context-length on line 149, but benchmark_lib.sh unconditionally unset MAX_MODEL_LEN for any caller under benchmarks/single_node/agentic/ (and check_env_vars deliberately omits it). Combined with set -euo pipefail at the top of this launcher, bash aborts with "MAX_MODEL_LEN: unbound variable" before the SGLang server is ever started — breaking all four dsv4-fp4-mi355x-sglang-agentic sweep configurations (TP8 ± DPA ± HiCache) 100% of the time. Fix by dropping the --context-length line entirely, matching every other agentic launcher in the repo (none pass --context-length/--max-model-len).
Extended reasoning...
The bug
Line 149 of benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh now builds SGLANG_CMD with:
--context-length "$MAX_MODEL_LEN"But MAX_MODEL_LEN is guaranteed to be unset at this point, and the script is running under set -euo pipefail (line 2), so parameter expansion under set -u aborts the script with MAX_MODEL_LEN: unbound variable before SGLANG_CMD is fully constructed. The SGLang server is never launched.
Why MAX_MODEL_LEN is unset
Line 15 sources benchmarks/benchmark_lib.sh. At module scope (benchmark_lib.sh:70-77) it deliberately unsets MAX_MODEL_LEN for any caller path matching */agentic/* (or SCENARIO_TYPE=agentic-coding, which also applies here per configs/amd-master.yaml):
# Agentic replays must use the model's native context limit. Ignore inherited
# workflow or shell overrides so neither the server nor AIPerf applies a cap.
_benchmark_caller="${BASH_SOURCE[1]:-}"
if [[ "$_benchmark_caller" == */agentic/* || ... || "${SCENARIO_TYPE:-}" == "agentic-coding" ]]; then
unset MAX_MODEL_LENThis launcher lives at benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh, so BASH_SOURCE[1] matches. The launcher's check_env_vars call at line 17 (MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION) intentionally omits MAX_MODEL_LEN — consistent with the library's stated policy that agentic replays use the model's native context. Nothing anywhere in the launcher re-assigns MAX_MODEL_LEN after the source. grep -n MAX_MODEL_LEN on this file returns exactly one hit: line 149.
How the sibling passed but this doesn't
The fixed-seq-len sibling (benchmarks/single_node/dsv4_fp4_mi355x_sglang.sh) also uses --context-length "$MAX_MODEL_LEN" and lists MAX_MODEL_LEN in its check_env_vars — that works because the sibling is not under agentic/, so benchmark_lib.sh never unsets it. The PR description notes the launcher was aligned with the fixed-seq-len sibling; unfortunately this one line's precondition (externally-supplied MAX_MODEL_LEN) does not survive the agentic path.
Step-by-step proof
- Sweep runner exports the standard agentic env (
MODEL,TP,CONC,KV_OFFLOADING,TOTAL_CPU_DRAM_GB,RESULT_DIR,DURATION,EP_SIZE,DP_ATTENTION,SCENARIO_TYPE=agentic-coding) and execs this launcher. - Line 2:
set -euo pipefail→ nounset active. - Line 15:
source .../benchmark_lib.sh.BASH_SOURCE[1]=benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh→ matches*/agentic/*→unset MAX_MODEL_LEN. - Line 17:
check_env_varssucceeds (it does not listMAX_MODEL_LEN). - Around line 149, bash evaluates the
SGLANG_CMD=( ... --context-length "$MAX_MODEL_LEN" ... )array literal."$MAX_MODEL_LEN"underset -u→dsv4_fp4_mi355x_sglang.sh: line 149: MAX_MODEL_LEN: unbound variable. Script exits. python3 -m sglang.launch_serveris never invoked;wait_for_server_readynever runs; the sweep row logs a launcher failure with an emptyserver.log.
Blast radius
configs/amd-master.yaml dsv4-fp4-mi355x-sglang-agentic search-space has exactly four rows (all tp: 8, conc-list: [48]): TP8 baseline, TP8 + HiCache, TP8 + DPA, TP8 + DPA + HiCache. Every one of them routes through this launcher, so the entire recipe is 100% broken by this PR.
Fix
Drop the --context-length "$MAX_MODEL_LEN" line. This matches every other agentic launcher in the repo — none of them pass --context-length or --max-model-len, deferring to the model's native context (which is exactly the policy the unset in benchmark_lib.sh is enforcing). If a cap is genuinely needed later, use --context-length "${MAX_MODEL_LEN:-<agentic-default>}".
| *) | ||
| echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: native, mooncake, lmcache)" >&2 | ||
| exit 1 |
There was a problem hiding this comment.
🟡 The default arm of the KV_OFFLOAD_BACKEND case at line 85 prints (expected: native, mooncake, lmcache), but the case above only accepts hicache — the listed values look copy-pasted from the sibling vLLM launchers (which do accept native/mooncake/lmcache) and would themselves fall into the same default arm here. Sweep configs only use hicache so CI never trips this branch, but anyone debugging by hand will be misdirected. Fix: change the list to (expected: hicache).
Extended reasoning...
What the bug is
The case "${KV_OFFLOAD_BACKEND:-}" block at lines 56-88 of benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh has exactly one accepting arm: hicache). Anything else falls into the default *) arm at lines 84-87, which emits:
Error: unsupported KV_OFFLOAD_BACKEND '<value>' (expected: native, mooncake, lmcache)
None of native, mooncake, or lmcache are recognized by this case statement — they are all backends supported by other launchers in the repo (native for kimik2.5-fp4-mi355x-vllm-agentic, mooncake for minimaxm3-fp8-mi300x-vllm-agentic, lmcache for the sibling dsv4-fp4-mi355x-vllm-agentic). The message is a copy-paste from those vLLM-family launchers.
Why the message is self-contradictory
A user who sees unsupported KV_OFFLOAD_BACKEND 'foo' and reads the parenthetical hint will pick one of native, mooncake, or lmcache — and hit the exact same default arm and the exact same message. In the case where they picked native, the output literally reads unsupported KV_OFFLOAD_BACKEND 'native' (expected: native, mooncake, lmcache), which is self-contradictory.
Step-by-step proof
- Operator invokes the launcher manually with
KV_OFFLOADING=dram KV_OFFLOAD_BACKEND=native(a natural attempt after seeing another dsv4 launcher uselmcache). agentic_kv_offload_enabledreturns true becauseKV_OFFLOADING=dram.- Bash enters the
caseat line 56 and evaluates the arms in order. The only arm ishicache). nativedoes not matchhicache), so control falls to*)at line 84.- Line 85 prints:
Error: unsupported KV_OFFLOAD_BACKEND 'native' (expected: native, mooncake, lmcache). - Line 86 exits 1.
- Operator, taking the message at face value, retries with
mooncakeorlmcache. Same path — steps 3-6 repeat.
Why existing code doesn't prevent it
The header comment at line 7 states "KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache", so the canonical documented value is hicache. The error text just wasn't updated to match when this launcher was adapted from the vLLM sibling.
Impact
Cosmetic misdirection in a diagnostic-only path. The four sweep entries in configs/amd-master.yaml for this recipe all pass kv-offload-backend: hicache, so CI never hits the default arm. The cost is a minute or two of wasted debug time for someone experimenting with a non-hicache backend by hand.
Fix
Change line 85's parenthetical to (expected: hicache):
echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: hicache)" >&2On the refutations
Both refutations are duplicate-management: bug_002 and bug_004 target the same defect at the same lines with the same fix, and the refuters explicitly note the finding itself is real — they only refuted one of the two to avoid posting the same comment twice. The synthesis agent merged them into a single report, which resolves that concern.
…config Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29076727890 |
…figs for initial validation Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29077198891 |
…unning-requests, re-enable all sweep configs - Fix attention-backend typo (compresseed -> compressed) - Remove duplicate CUDA_GRAPH_MAX_BS block - Set MAX_RUNNING_REQUESTS to 2*CONC for subagent fan-out headroom - Use MEM_FRACTION_STATIC=0.90 and swa-full-tokens-ratio=0.10 - Add METRICS_ARGS array, guard capture_cache_metrics on metrics enabled - Remove duplicate SGLANG_ENABLE_UNIFIED_RADIX_TREE export from hicache block - Uncomment all 4 sweep configs (TP8 +/- DPA +/- HiCache) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29081203249 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29081361033 |
…ust sweep concurrency hicache-ratio=8 tried to pin ~495 GB via cudaHostRegister, exceeding ROCm's per-process pinnable memory limit. Lower to ratio=4 and sweep graduated concurrency levels across configs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29094196523 |
…concurrency to 96 Device KV utilization peaked at only 27% even at conc=48, and non-DPA configs underperformed DPA in both throughput and completed requests. Drop non-DPA sweep entries and extend DPA concurrency range to [16..96] (none) and [48..96] (hicache) to find the saturation point. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29108708524 |
1 participant
Summary
dsv4-fp4-mi355x-sglang-agenticimage tolmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710dsv4attention backend,fp8_e4m3KV cache,--disable-radix-cache,--cuda-graph-max-bs, DP-attention exports,--enable-two-batch-overlap)SGLANG_CMDarray pattern with command logging and env-var dumpcapture_cache_metricsfor pre/post-benchmark cache statsconc=48across TP8 +/- DPA +/- HiCacheTest plan
dsv4-fp4-mi355x-sglang-agenticfull sweep on MI355X cluster🤖 Generated with Claude Code