[deepseek_r1] General PP enabling

[tvoas]

• Enable PP solution with full support for DeepSeek R1 execution with PP>1.
• Requires 1.21.0 or newer. Does not support 1.20.1 or older.
• Implementation mirrors Implement Pipeline Parallelism support for HPU. #1000 as closely as possible while ensuring DeepSeek R1 functions fully.
• Adds a benchmark script for sweeping various configs automatically. This can be removed if you feel it shouldnt merge to deepseek_r1 branch.

Additional validation is being done by yabai.hu@intel.com.

@czhu15 youlei.yang@intel.com please help start the review in the meantime.

[czhu15]

why we need set "VLLM_USE_FP8_MATMUL" by application? shouldn't INC handle it?
And what's VLLM_USE_SINGLE_TENSOR_CACHE for? Can't find it in the vllm code...

[tvoas]

Good cache. This was a leftover from Liu, Yi's FP8 matmul implementation on the PRC local version of this branch. I see that the implementation that is provided on deepseek_r1 branch is different, though and no longer uses VLLM_USE_SINGLE_TENSOR_CACHE.

We still need VLLM_USE_FP8_MATMUL though, right? This environment variable should be set for improved performance according to @yiliu30.

[yiliu30]

Yes, the VLLM_USE_SINGLE_TENSOR_CACHE flag should be removed, as we have been using one tensor for KVCache since #977.

We still need VLLM_USE_FP8_MATMUL for FP8 Q@K and FP8 A@V, since the current path does not use the INC to replace the Matmul with PatchedMatmul. Instead, it manually patches Q@K and A@V and use 1.0 as the scaling. Please refer to #977 for more details.

cc @xuechendi