[gaudi] HuggingFaceM4/idefics2-8b issue fix

batch.prefill_cache_indices is reset in generate_token instead of forward, so that position_id could be updated correctly

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

backends/gaudi/server/text_generation_server/models/custom_modeling/flash_mistral_modeling.py

@@ -111,7 +111,7 @@ class MistralAttention(torch.nn.Module):
         )
         self.num_heads = config.num_attention_heads
         self.hidden_size = config.hidden_size
-        if hasattr(config, "head_dim"):
+        if getattr(config, "head_dim", None) is not None:
             self.head_size = config.head_dim
         else:
             self.head_size = self.hidden_size // self.num_heads

backends/gaudi/server/text_generation_server/models/flash_vlm_causal_lm.py

@@ -1050,8 +1050,6 @@ class FlashVlmCausalLM(FlashCausalLM):
             attention_mask=attention_mask_forward,
             **kwargs,
         )
-        if batch.prefill_cache_indices is not None:
-            batch.prefill_cache_indices = None
         batch.image_grid_thw = None
         batch.free_encoder_cache()
         return logits, speculative_logits