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

[EdisonSu768]

Adds the validated-model page and deployment manifest for Gemma-4-31B-it on Ascend 910B3.

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: vLLM-Ascend commit b481d79c7de49ab285e61cec689172931c124429; public ARM64 image pinned by manifest sha256:fa246685a9e0b9ea63f242be634dcc7616060f6f7dc9ca7018fa06c1657940da (CANN 9.0.1).
• Prefill: TP4, max-num-batched-tokens=8192, PIECEWISE ACLGraph.
• Decode: TP4, max-num-batched-tokens=128, FULL_DECODE_ONLY graph capture.
• P/D transport: MooncakeConnectorV1 with Ascend Direct KV transfer.
• Ingress under test: MaaS /v1/chat/completions; API key injected from a Kubernetes Secret.

AIPerf results

AIPerf 0.7.0, closed loop, concurrency 8/16/32, 480 requests per tier. All six tiers completed through MaaS with 0 errors; all four serving Pods remained at 0 restarts.

Scenario	conc32 TTFT	conc32 ITL	conc32 req/s	conc32 total tok/s
1: 8,013 input / 128 output	26.843 s	29.36 ms	1.01	8,257.80
2: 17,933 average input / 128 output	67.295 s	28.85 ms	0.44	7,862.21

Throughput and ITL are flat from concurrency 8 onward. The service remains Prefill-capacity-bound at roughly 8k input-dominated tokens/s; higher concurrency increases TTFT without increasing throughput. Compared with the 2026-08-03 MaaS run on the older engine/configuration, total TPS is about 0.9%-1.0% lower, average TTFT about 1.1%-2.4% higher, and ITL about 0.3%-0.6% lower: effectively unchanged rather than a material speedup.

capacity14 Agent results

Pi 0.80.6 and OpenCode 1.17.18 ran the fixed 14-task Terminal-Bench capacity14 set through MaaS with reasoning/thinking enabled. Concurrency 32 uses three attempts per task (42 trials).

Agent	Concurrency	Completed	Passed	Pass rate	Timeouts	Correct tasks/hr
Pi	1	14 / 14	5 / 14	35.7%	0	3.60
Pi	8	14 / 14	7 / 14	50.0%	1	25.89
Pi	32	42 / 42	20 / 42	47.6%	15	17.40
OpenCode	1	14 / 14	7 / 14	50.0%	0	4.41
OpenCode	8	14 / 14	6 / 14	42.9%	0	22.47
OpenCode	32	42 / 42	15 / 42	35.7%	16	17.79

Pi/OpenCode c32 Pass@2 is 54.8%/42.9%. Both agents achieve their best correct-task throughput at concurrency 8; c32 adds many Agent timeouts without improving quality or useful throughput. The serving Pods stayed Ready with 0 restarts and no serving 5xx/traceback, so these are Agent/task-level timeouts rather than an engine crash. mini7 is retained as smoke evidence only and is not published as the formal Agent score.

Deployment corrections included

• Adds --reasoning-parser gemma4 and keeps the Gemma tool-call parser.
• Replaces Prefill eager mode with PIECEWISE ACLGraph.
• Pins the exact public ARM64 image manifest used for validation.
• Uses the validated Ascend Direct RoCE/HCCN mount and non-conflicting KV ports.
• Keeps the single-pair scheduler without the tokenizer sidecar and sets its metrics refresh interval to 15 seconds.

Validation

• yarn lint: 0 errors, 0 warnings.
• YAML stream parsed successfully (4 documents).
• git diff --check passed.

[cloudflare-workers-and-pages]

Deploying alauda-ai with    Cloudflare Pages

Latest commit:	48f3830
Status:	✅  Deploy successful!
Preview URL:	https://292a6359.alauda-ai.pages.dev
Branch Preview URL:	https://docs-gemma4-31b-it-pd-aiperf.alauda-ai.pages.dev

View logs