Skip to content

docs(inference_guide): validate Gemma-4-31B-it P4+D4 on Ascend 910B3 - #298

Open
EdisonSu768 wants to merge 2 commits into
masterfrom
docs/gemma4-31b-it-pd-aiperf
Open

docs(inference_guide): validate Gemma-4-31B-it P4+D4 on Ascend 910B3#298
EdisonSu768 wants to merge 2 commits into
masterfrom
docs/gemma4-31b-it-pd-aiperf

Conversation

@EdisonSu768

@EdisonSu768 EdisonSu768 commented Aug 2, 2026

Copy link
Copy Markdown
Member

Adds a validated inference-guide page and deployment manifest for Gemma-4-31B-it. Deployment and benchmark evidence come from model-auto commits dcad16c and 966538d on the Guian Ascend 910B3 cluster.

Validated stack

  • Model: google/gemma-4-31B-it, BF16 text-generation path.
  • Topology: P4+D4, 8 x Ascend 910B3 64 GB across two nodes.
  • Engine: quay.io/ascend/vllm-ascend:nightly-main-openeuler-0721 (vLLM 0.25.1, CANN 9.0.1).
  • P/D transport: MooncakeConnectorV1 with Ascend Direct KV transfer.
  • Ingress under test: direct P/D proxy; MaaS still selected the old aggregate TP4 service.

AIPerf results

AIPerf 0.7.0, closed loop, concurrency 8/16/32, 480 requests per tier. All six tiers completed with 0 errors.

Scenario conc32 TTFT conc32 ITL conc32 req/s conc32 total tok/s
1: 8,013 input / 128 output 26.458 s 29.09 ms 1.03 8,372.31
2: 17,933 average input / 128 output 66.372 s 28.54 ms 0.44 7,970.40

Throughput and ITL are flat from concurrency 8 onward. The service is Prefill-capacity-bound at roughly 8k input-dominated tokens/s; higher concurrency increases TTFT without increasing throughput.

Scheduler correction

The delivered manifest removes the tokenizer plugin and sidecar. This deployment has one fixed P/D pair and no prefix-cache producer, so tokenizer-based routing adds no choice. Hermes Router 26.6.0-rc.3 used a 5-second initialization deadline while the Gemma tokenizer cold-loaded in 17.59 seconds, causing endless retries. The live scheduler rolled from 2/2 to 1/1 Ready without restarting the 11-day-old Prefill/Decode Pods, and direct P/D smoke remained HTTP 200.

Validation

  • yarn lint: 0 errors, 0 warnings.
  • YAML stream parsed successfully.
  • git diff --check passed.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Aug 2, 2026

Copy link
Copy Markdown

Deploying alauda-ai with  Cloudflare Pages  Cloudflare Pages

Latest commit: 6885cda
Status: ✅  Deploy successful!
Preview URL: https://8b0fe7fa.alauda-ai.pages.dev
Branch Preview URL: https://docs-gemma4-31b-it-pd-aiperf.alauda-ai.pages.dev

View logs

@EdisonSu768 EdisonSu768 changed the title docs(inference_guide): validate Gemma-4-31B-it on Ascend 910B3 P4+D4 docs(inference_guide): validate Gemma-4-31B-it P4+D4 on Ascend 910B3 Aug 2, 2026
@EdisonSu768
EdisonSu768 force-pushed the docs/gemma4-31b-it-pd-aiperf branch from a2b2013 to 6885cda Compare August 3, 2026 12:55
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant