add encoder cache free

server/text_generation_server/models/flash_causal_lm.py

@@ -1899,7 +1899,9 @@ class FlashCausalLM(Model):
             batch.prepare_for_prefill()
 
         self.get_input_embeddings(batch)
+        from pdb import set_trace
 
+        set_trace()
         prefill_logprobs = batch.prefill_next_token_indices is not None
 
         # Update adapter indices for speculative tokens (if present)

server/text_generation_server/models/transformers_flash_vlm.py

@@ -304,6 +304,7 @@ class TransformersFlashVlmCausalLM(VlmCausalLM):
             device=device,
             rank=rank,
             world_size=world_size,
+            support_chunking=True,
         )
 
         # Monkey patch of `self.model.forward` to match `FlashCausalLM`. It avoids duplicating a lot of code

server/text_generation_server/models/vlm_causal_lm.py

@@ -499,6 +499,7 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
                 if image_id not in self.encoder_cache[i]:
                     self.scheduled_image_input.append((i, image_position))
                     scheduled_image_pixel_values.append(pixel_values)
+                    self.image_inputs[i][j] = None
 
         if self.has_image and len(scheduled_image_pixel_values):
             self.pixel_values = torch.cat(
@@ -590,6 +591,37 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
 
         return torch.cat(mm_embeds, dim=0).to(device)
 
+    def free_encoder_cache(self):
+        for i, (
+            r,
+            cache_length,
+            input_length,
+            request_prefilling,
+        ) in enumerate(
+            zip(
+                self.requests,
+                self.cache_lengths,
+                self.input_lengths,
+                self.prefilling_mask,
+            )
+        ):
+            if not request_prefilling or self.image_positions[i] is None:
+                continue
+
+            for j, image_position in enumerate(self.image_positions[i]):
+                image_id = image_position.id
+
+                start_pos = image_position.offset
+                length = image_position.length
+
+                cache_length = cache_length + input_length
+                if start_pos >= cache_length:
+                    # No encoder input required at this step
+                    break
+
+                if start_pos + length <= cache_length:
+                    self.encoder_cache[i][image_id] = None
+
 
 class VlmCausalLM(FlashCausalLM):
     def __init__(
@@ -831,4 +863,6 @@ class VlmCausalLM(FlashCausalLM):
             else None
         )
         logits = cuda_graph["logits"][:bs]
+
+        batch.free_encoder_cache()
         return logits, speculative_logits