replaced apply_rotary_emb_torch() with rotary_embedding imp

[slokesha]

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

Purpose

Test Plan

Test Result

[libinta]

try hpu version def apply_rotary_pos_emb_hpu(
q: torch.Tensor,
k: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
# Determine rotary dimension from cos/sin shape
ro_dim = cos.shape[-1] * 2

# Split into rotated and pass-through parts  
q_rot = q[..., :ro_dim]  
q_pass = q[..., ro_dim:]  
k_rot = k[..., :ro_dim]  
k_pass = k[..., ro_dim:]  
  
# Prepare cos/sin (remove the chunking)  
cos_full = cos  
sin_full = sin  
from habana_frameworks.torch.hpex.kernels import apply_rotary_pos_emb
q_rot = apply_rotary_emb_torch(q_rot.float(), cos_full.float(), sin_full.float()).type_as(q)  
 k_rot = apply_rotary_emb_torch(k_rot.float(), cos_full.float(), sin_full.float()).type_as(k)  
  
 # Concatenate rotated and pass-through parts  
 q_embed = torch.cat([q_rot, q_pass], dim=-1)  
 k_embed = torch.cat([k_rot, k_pass], dim=-1)  
  
return q_embed, k_embed

[slokesha]

Updated PR with suggested code. Had to make changes in apply_rotary_emb_torch to support full cos and sin instead of chunked. Also you suggested from to import apply_rotary_pos_emb from habana_frameworks.torch.hpex.kernels - But was not used.

If we wanted :
from habana_frameworks.torch.hpex.kernels import apply_rotary_pos_emb
q_rot = apply_rotary_pos_emb(q_rot.float(), cos_full.float(),
sin_full.float()).type_as(q)
k_rot = apply_rotary_pos_emb(k_rot.float(), cos_full.float(),
sin_full.float()).type_as(k)

We get this error: File "/usr/local/lib/python3.12/dist-packages/vllm-0.6.3.dev5516+geb01e1dc5.gaudi122-py3.12.egg/vllm/model_executor/models/siglip2navit.py", line 294, in forward
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Graph compile failed. Recipe: .graph_dumps/HabanaFusedOpLazy_21_21, synStatus=synStatus 26 [Generic failure].

With the current PR:

Metric	vs H100	vs GT	All-core Query Time [s]
Ovis (no Media pipeline)	98.08%	94.23%	1st run -63.03 & 2nd run-47.67