[INTEL_HPU] supported ERNIE-4.5-21B-A3B-Thinking mold

[FocusLuo]

ERNIE-4.5-21B-A3B-Thinking needs to use DefaultModelLoaderV1 mode

reference command line:
ENABLE_V1_KVCACHE_SCHEDULER=1 FD_ENC_DEC_BLOCK_NUM=8 HPU_PERF_BREAKDOWN_SYNC_MODE=1 \ HPU_WARMUP_BUCKET=0 MAX_PREFILL_NUM=1 FD_ATTENTION_BACKEND=HPU_ATTN \ python -m fastdeploy.entrypoints.openai.api_server --model \ ./models--baidu--ERNIE-4.5-21B-A3B-Thinking/snapshots/4341bb42644d5422859509fa25d41544c57181f8/ \ --port 8388 --engine-worker-queue-port 8302 --metrics-port 8301 \ --cache-queue-port 8303 --max-model-len 16384 --tensor-parallel-size 1 \ --load-choices "default_v1" --num-gpu-blocks-override 5000 --kv-cache-ratio 0.5 \ --max-num-seqs 128 --block-size 64 --no-enable-prefix-caching \ --graph-optimization-config '{"use_cudagraph":false}'

python bench_gsm8k.py --data-path ./test.jsonl --port 8388 --num-shots 5 --num-questions 1319 --parallel 1 --result-file test_accuracy_parallel_1.json

Motivation

supported ERNIE-4.5-21B-A3B-Thinking mold

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

enabled DefaultModelLoaderV1 on Intel HPU

Usage or Command

ENABLE_V1_KVCACHE_SCHEDULER=1 FD_ENC_DEC_BLOCK_NUM=8 HPU_PERF_BREAKDOWN_SYNC_MODE=1 HPU_WARMUP_BUCKET=0 MAX_PREFILL_NUM=1 FD_ATTENTION_BACKEND=HPU_ATTN python -m fastdeploy.entrypoints.openai.api_server --model /mnt/disk3/ernie_opensource/hub/models--baidu--ERNIE-4.5-21B-A3B-Thinking/snapshots/4341bb42644d5422859509fa25d41544c57181f8/ --port 8388 --engine-worker-queue-port 8302 --metrics-port 8301 --cache-queue-port 8303 --max-model-len 16384 --tensor-parallel-size 1 --load-choices "default_v1" --num-gpu-blocks-override 5000 --kv-cache-ratio 0.5 --max-num-seqs 128 --block-size 64 --no-enable-prefix-caching --graph-optimization-config '{"use_cudagraph":false}'

Accuracy Tests

On Intel HPU
{"task": "gsm8k", "backend": "paddlepaddle", "num_gpus": 1, "latency": 21223.108, "accuracy": 0.774, "num_requests": 1319, "other": {"num_questions": 1319, "parallel": 1}}
{"task": "gsm8k", "backend": "paddlepaddle", "num_gpus": 1, "latency": 3019.105, "accuracy": 0.754, "num_requests": 1319, "other": {"num_questions": 1319, "parallel": 16}}

On NV H20
ENABLE_V1_KVCACHE_SCHEDULER=1 python -m fastdeploy.entrypoints.openai.api_server --model ./models--baidu--ERNIE-4.5-21B-A3B-Thinking/snapshots/4341bb42644d5422859509fa25d41544c57181f8/ --port 8388 --engine-worker-queue-port 8302 --metrics-port 8301 \ --cache-queue-port 8303 --tensor-parallel-size 1 --max-model-len 16384 --max-num-seqs 128 --block-size 64 --kv-cache-ratio 0.5 --num-gpu-blocks-override 6000 --graph-optimization-config '{"use_cudagraph":false}' --no-enable-prefix-caching --load-choices "default_v1"
{"task": "gsm8k", "backend": "paddlepaddle", "num_gpus": 1, "latency": 18805.898, "accuracy": 0.763, "num_requests": 1319, "other": {"num_questions": 1319, "parallel": 1}}

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@adb91dc). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/model_executor/utils.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5891   +/-   ##
==========================================
  Coverage           ?   66.59%           
==========================================
  Files              ?      347           
  Lines              ?    44467           
  Branches           ?     6835           
==========================================
  Hits               ?    29614           
  Misses             ?    12668           
  Partials           ?     2185           

Flag	Coverage Δ
GPU	66.59% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[FocusLuo]

@LeoZhao-Intel @yanfeich @JianyuLi01 @feiwan1 @fmiao2372 could you help to give a review?

[LeoZhao-Intel]

suggest to add "intel hpu" into _err_msg.

[FocusLuo]

thanks review, updated and added,

[zoooo0820]

LGTM