[Perf][ROCm] Use AITER gluon sparse MLA for rope-free BF16 (GLM-5.3-Flash)

[simondanielsson]

Purpose

Awaiting; AITER bump

GLM-5.3-Flash doesn't use RoPE and falls back to the default triton impl for BF16 KV MLA.

ROCm/aiter#4919 added a gfx950 gluon kernel tailored for BF16 KV non-rope models. We add it here.

Perf gain: TODO

Fixes #57588.

Test Plan

Run GLM-5.3-Flash on MI350 (gfx950), TP8, BF16 KV cache.

1. GSM8k vs main

Expand for details

VLLM_ROCM_USE_AITER=1 vllm serve zai-org/GLM-5.3-Flash \
    --tensor-parallel-size 8 \
    --max-num-seqs 512 \
    --gpu-memory-utilization 0.8 \
    --attention-backend ROCM_AITER_MLA_SPARSE \
    --max-num-batched-tokens 16384

lm_eval --model local-completions --tasks gsm8k \
    --model_args model=zai-org/GLM-5.3-Flash,base_url=http://127.0.0.1:8000/v1/completions \
    --num_fewshot 5

2. bench serve 1k/1k and 8k/1k vs main

Both prefill and decode move to the new kernel, so TTFT and TPOT are both
reported.

Expand for details

vllm bench serve \
    --backend vllm \
    --model zai-org/GLM-5.3-Flash \
    --dataset-name random \
    --random-input-len 1000 \
    --random-output-len 1000 \
    --ignore-eos \
    --seed 5678 \
    --max-concurrency $CONC \
    --num-prompts $((4 * CONC)) --num-warmups $CONC

3. Non-regression on the separated-rope path

DeepSeek-V4-Flash on the same backend, to confirm the ASM and Triton routes are
unchanged.

Expand for details

VLLM_ROCM_USE_AITER=1 vllm serve deepseek-ai/DeepSeek-V4-Flash \
    --tensor-parallel-size 8 --kv-cache-dtype fp8_ds_mla \
    --attention-backend ROCM_AITER_MLA_SPARSE

Test Result

1. GSM8k

TODO

2. 1k/1k

TODO (P50s)

Concurrency	Variant	QPS	TTFT (ms)	TPOT (ms)	% TPOT improved

3. Non-regression on the separated-rope path

TODO

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[cagrikymk]

@simondanielsson Hello, I am the original author of the MLA kernel on the aiter side.

I have a draft ready here: #53492 I was waiting for my PR to be merged to mark the PR ready.

I am planning to extend this API for dsv4, dsv4-flash, GLM-5.x and GLM 5.3-flash. So, could we prioritize my PR over this?

[simondanielsson]

@simondanielsson Hello, I am the original author of the MLA kernel on the aiter side.

I have a draft ready here: #53492 I was waiting for my PR to be merged to mark the PR ready.

I am planning to extend this API for dsv4, dsv4-flash, GLM-5.x and GLM 5.3-flash. So, could we prioritize my PR over this?

@cagrikymk Thanks for letting me know! Indeed let's prio yours and I'll close this one.

I think the tests I added in this PR we can probably port to yours as well, wdyt?