[AMD][AgentX] DeepSeek-V4 MI355X agentic disaggregated benchmarking#2170
[AMD][AgentX] DeepSeek-V4 MI355X agentic disaggregated benchmarking#2170ichbinblau wants to merge 31 commits into
Conversation
ichbinblau
requested review from
1am9trash,
billishyahao,
chunfangamd,
seungrokj and
yctseng0211
as code owners
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29140974686 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29141073508 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29141213005 |
billishyahao
left a comment
billishyahao
left a comment
There was a problem hiding this comment.
As a PR reviewer and CODEOWNER, I have reviewed this and have:
- Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
- Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse. Approve for day 0
- Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29141213005
- Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/29141213005
- Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
- Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; FLOPs at lower precisions is fine, given that the config passes private evals. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
- If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
- If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
- Verified that the single-node recipes are similar to the official vLLM recipes and/or theSGLang cookbook:
- If they are not, I have verified that a PR has been opened in vLLM recipe repo or SGLang repo and linked it below in the additional detail section:
- If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
Additional detail section:
- insert any additional info here
This is the re-submission of #2127
Signed: @billishyahao
❌❌❌ REJECTED ❌❌❌@billishyahao blocking: no evals ran on the signed-off commit, no authorized ✅ Check 0 (CODEOWNER): PASS — you own |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29222483374 |
| # DSV4 uses page-size 256 (set in models.yaml); HiCache must match. | ||
| export HICACHE_PAGE_SIZE="${HICACHE_PAGE_SIZE:-256}" | ||
| # HiCache ratio (host pool = ratio * GPU KV pool). Default derived in server_sglang.sh. | ||
| export HICACHE_RATIO="${HICACHE_RATIO:-}" |
There was a problem hiding this comment.
@cquil11 plz check how this var works and if it follows the rule of ALLOWED_CPU_MEM = amount_of_cpu_mem * (# of gpus used per node / total GPUs per node)
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29223014792 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29229732720 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29235085283 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29237371210 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29237371210 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29265000316 |
ichbinblau
force-pushed
the
amd/agentx-v1.0-yanfei
branch
from
12b3333 to
94883e7
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29307961980 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29307996568 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29308507513 |
patch_decode_tp_queue_agree()'s invocation was already commented out (dead code with no runtime effect), and the reference-only patches/decode_tp_queue_agree.patch it mirrored had no other callers. Drop both plus the README bullet pointing at it. Co-authored-by: Cursor <cursoragent@cursor.com>
…desync patch install_transformers_glm5() was gated on an exact MODEL_NAME == "GLM-5-FP8" match; broaden to any model name containing "GLM" so other GLM variants pick up the same glm_moe_dsa transformers fix. Also disable patch_disagg_prefill_bootstrap_desync's invocation (commented out, matching the already-disabled decode_tp_queue_agree pattern removed earlier). Co-authored-by: Cursor <cursoragent@cursor.com>
…2147) * agentic: add node-0 sibling benchmark-client container for DSv4 sweeps Port the "same-node sibling container" client mode from ROCm/InferenceY: when CLIENT_IMAGE is set (and no CLIENT_NODES), node 0 launches the aiperf trace replay in its own pre-baked container via the host docker socket, instead of rebuilding the aiperf venv inside the server container and running it co-located. This keeps the client's CPU-heavy tokenize/aggregate work off the sglang scheduler + router, which inflates TTFT/E2E and lowers throughput under agentic concurrency. - server_sglang.sh: add IS_AGENTIC_RUN and a CLIENT_IMAGE sibling-container branch that writes client.env and docker-runs the client against the local router (--network host). - job.slurm: define CLIENT_CONT_NAME; when CLIENT_IMAGE is set, mount the host docker socket + CLI into the server container, forward HOST_REPO_DIR/HOST_MODEL_DIR/HOST_BENCH_LOGS/CLIENT_CONT_NAME, pre-pull the client image, and clean up the client container on teardown. - amd-master.yaml: enable the sibling client on dsv4-fp4-mi355x-sglang-disagg-agentic-hicache via CLIENT_IMAGE. The separate-client-NODE mode is intentionally not ported. * agentic: use server image for sibling client so upstream CI can pull it The pre-baked rocm/pytorch-private aiperf client image is not pullable by upstream CI runners. Reuse the (public) server image as CLIENT_IMAGE and build aiperf on the fly from /workspace/utils/aiperf, matching the co-located path. Gate the pre-baked-venv env (AIPERF_USE_PREBUILT / AIPERF_VENV) behind an optional CLIENT_AIPERF_VENV so a real pre-baked client image can still opt in.
Signed-off-by: Theresa Shan <theresa.shan@amd.com>
Update the DSv4 MI355X agentic perf-changelog entry to point at PR 2170 with an accurate description, and move hf_dir into the DeepSeek-V4-Pro YAML block so the SGLang model path extractor can read it. 中文：修复 PR 2170 的 review 反馈：将 DSv4 MI355X agentic 的 perf-changelog 条目改为指向 PR 2170 并更新为准确描述，同时把 hf_dir 移入 DeepSeek-V4-Pro YAML 配置块，确保 SGLang 模型路径提取逻辑可以读取。
The disaggregation-prefill bootstrap-desync patch was already disabled (commented out); drop the function, its comment block, and the stale invocation now that the fix is carried upstream in the pinned image. Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Theresa Shan <theresa.shan@amd.com>
Resolve the models.yaml entry per run type: agentic runs (IS_AGENTIC) use the '<model>-AgentX' recipe, non-agentic disaggregated runs use '<model>-DI', falling back to the bare model name when no variant key exists. Apply the same logic in both server_sglang.sh (lookup) and job.slurm (validation guard) so a bare MODEL_NAME (e.g. DeepSeek-V4-Pro) correctly maps to the suffixed key. Add DeepSeek-V4-Pro-DI and rename the agentic recipe to DeepSeek-V4-Pro-AgentX. Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Theresa Shan <theresa.shan@amd.com>
- server_sglang.sh: lower default HICACHE_RATIO 24->5 (cut host-pool pin/registration startup cost) and raise router health barrier timeout 1800->3000 so a slow server bring-up no longer trips the cross-node router wait first. - amd-master.yaml: restore conc 32 on the non-DP-attn (TP8) point and comment out the DP-attn c64/128 sweep block. - models.yaml: drop explicit max_total_tokens override for the AgentX dp variant (let it derive from context_length). Co-authored-by: Cursor <cursoragent@cursor.com>
Signed-off-by: Theresa Shan <theresa.shan@amd.com>
perf-changelog.yaml must only gain entries; re-adding the upstream qwen3.5-fp8-gb300-dynamo-sglang block fixes process_changelog.py. Co-authored-by: Cursor <cursoragent@cursor.com>
Use kv-offload-backend object form and add kv-p2p-transfer: mori for the disaggregated dsv4-fp4-mi355x-sglang-disagg-agentic-hicache entry. Co-authored-by: Cursor <cursoragent@cursor.com>
Override models.yaml --disaggregation-transfer-backend when kv-p2p-transfer from amd-master.yaml is set via CI, and forward KV_P2P_TRANSFER to the agentic client env for result metadata. Co-authored-by: Cursor <cursoragent@cursor.com>
Let SGLang auto-size the prefill KV pool instead of pinning 1M tokens. Co-authored-by: Cursor <cursoragent@cursor.com>
Give prefill/decode workers more time to wait for the node-0 proxy during long agentic server startup. Co-authored-by: Cursor <cursoragent@cursor.com>
Add kv-p2p-transfer to disagg multinode test configs and use KVOffloadBackendMetadata dict form for kv-offload-backend fields. Co-authored-by: Cursor <cursoragent@cursor.com>
ichbinblau
force-pushed
the
amd/agentx-v1.0-yanfei
branch
from
f8dcfba to
6344aaa
Compare
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29310514664 |
Extract SYNC_BARRIER_TIMEOUT (3000s for DeepSeek-V4, 1800s otherwise) and apply it to server-up, router readiness, and prefill/decode proxy barriers. Also lower the default HICACHE_RATIO from 16 to 5. Co-authored-by: Cursor <cursoragent@cursor.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29310856104 |
Pass KV_OFFLOAD_BACKEND_METADATA into server containers via job.slurm and forward it to sibling client containers in server_sglang client.env so process_agentic_result can write aggregate JSON. Derive KV_OFFLOAD_BACKEND from metadata in the DSv4 agentic recipe when unset. Co-authored-by: Cursor <cursoragent@cursor.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29314804396 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29327827754 |
CI already sets KV_OFFLOAD_BACKEND from the workflow; keep the job.slurm and server_sglang metadata forwarding fixes without the redundant recipe logic. Co-authored-by: Cursor <cursoragent@cursor.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=29327827754 |
toJson() pretty-prints with embedded newlines, so the unquoted expansion in
job.slurm's DOCKER_ENV_COMMON was word-split across the docker run command,
corrupting its argument list ("docker: invalid reference format", exit 125).
Quote the value so the whole JSON stays one token.
Co-authored-by: Cursor <cursoragent@cursor.com>
7 participants
Summary
Adds end-to-end support for DeepSeek-V4 (DSv4) agentic-coding benchmarks on MI355X in the SGLang disaggregated (prefill/decode) path, plus the multi-node sweep plumbing, result aggregation, and CI fixes needed to make those runs green and ingestible.