Support Qwen3 Embedding & Reranker on Gaudi for aice/v1.21.0 branch

[gyou2021]

1. Support Qwen3 Embedding & Reranker on Gaudi
2. Enabled HPU lazy mode of Roberta model on Gaudi.

Code modifications on HPU:

1. Cherry-pick commit 3952731 of vllm-project on GPU
2. Fixed the bug in hpu_model_runner.py to enable correct call.
3. Updated pooler.py of aice/v1.21.0 branch of vllm-fork based on that of the main branch of vllm-project and fixed the score bug.
4. Removed assert in roberta.py to enable hpu_graph mode.

example:
cd examples/offline_inference
PT_HPU_LAZY_MODE=1 VLLM_SKIP_WARMUP=true python qwen3_reranker.py

Supported on aice/v1.21.0.

[czhu15]

"Currently, not supported on aice/v1.21.0 yet."
Is this feature/PR ready for review/merge?

[czhu15]

Will be good to provide the link the original PR and highlight the changes on HPU.
This will make our future debug easy.
Thanks!

[ranzhejiang]

I notice that you use _hpu_merge_multimodal_embeddings, but #1436 has remove this function and using another way, can you use the same approach?

[gyou2021]

"Currently, not supported on aice/v1.21.0 yet." Is this feature/PR ready for review/merge?

Yes.

[gyou2021]

I notice that you use _hpu_merge_multimodal_embeddings, but #1436 has remove this function and using another way, can you use the same approach?

The code has been rebased to the latest aice/v1.21.0, including changes made by #1436.

[czhu15]

@gyou2021
Will be good to provide the link the original PR and highlight the changes on HPU.
This will make our future debug easy.
Can you add some description on the commit message?

[gyou2021]

@gyou2021 Will be good to provide the link the original PR and highlight the changes on HPU. This will make our future debug easy. Can you add some description on the commit message?
the original PR on vllm-project:
vllm-project@3952731

@gyou2021 Will be good to provide the link the original PR and highlight the changes on HPU. This will make our future debug easy. Can you add some description on the commit message?

Sure. Added the original PR and highlight the changes on HPU in the above PR descriptions.