Fix long contexts in LoRA (#624)

#566 breaks long-contexts +
LoRA flow.

This assumes caching sin-cos buffer for first decoder layer is
sufficient to handle all cases, which is not the applicable for
long-context + LoRA.

This PR ignores `_prepare_cos_sin` call prior to HpuModelAdapter forward
in long-context + LoRA flow.

Author: scsudhak-intel

tests/lora/conftest.py

@@ -64,6 +64,7 @@ def cleanup_fixture(should_do_global_cleanup_after_test: bool):
 
 @pytest.fixture
 def dist_init():
+    import habana_frameworks.torch.hpu  # noqa: F401
     temp_file = tempfile.mkstemp()[1]
     backend_type = "hccl" if current_platform.is_hpu() else "nccl"
     init_distributed_environment(

vllm/lora/punica_wrapper/utils.py

@@ -2,6 +2,8 @@
 
 import torch
 
+from vllm.platforms import current_platform
+
 if TYPE_CHECKING:
     # avoid circuit import
     from vllm.lora.layers import LoRAMapping
@@ -86,10 +88,14 @@ def convert_mapping(
     embedding_indices = index_mapping_indices.copy()
     lora_indices = index_mapping_indices.copy()
     long_lora_offsets: Optional[torch.Tensor] = None
+
     if long_lora_context:
-        long_lora_offsets = torch.zeros(len(index_mapping_indices),
-                                        device=device,
-                                        dtype=torch.long)
+        if current_platform.is_hpu():
+            long_lora_offsets_list: List[int] = []
+        else:
+            long_lora_offsets = torch.zeros(len(index_mapping_indices),
+                                            device=device,
+                                            dtype=torch.long)
     prompt_mapping: List[int] = [
         lora_index_to_id.index(x) if x > 0 else -1
         for x in mapping.prompt_mapping
@@ -102,10 +108,18 @@ def convert_mapping(
         embedding_indices[i] = lora_idx if index_mapping_indices[i] > 0 else 0
         lora_indices[i] = lora_idx
         if long_lora_context:
-            assert long_lora_offsets is not None
             lora_offset: int = long_lora_context.offsets_by_lora_id.get(
                 index_mapping_indices[i], 0)
-            long_lora_offsets[i] = lora_offset
+            if current_platform.is_hpu():
+                long_lora_offsets_list.append(lora_offset)
+            else:
+                assert long_lora_offsets is not None
+                long_lora_offsets[i] = lora_offset
+
+    if long_lora_context and current_platform.is_hpu():
+        long_lora_offsets = torch.tensor(long_lora_offsets_list,
+                                         device=device,
+                                         dtype=torch.long)
 
     indices_list: List[Union[List[int], torch.Tensor]] = [
         index_mapping_indices,

vllm/model_executor/layers/rotary_embedding.py

@@ -232,7 +232,10 @@ def forward_hpu(
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         from habana_frameworks.torch.hpex.kernels import (
             RotaryPosEmbeddingMode, apply_rotary_pos_emb)
-        if self.sin is None:
+
+        # Prepare cos-sin caches for long-context + LoRA with offsets for every
+        # forward, since the offset information wasn't available previously
+        if hasattr(self, "scaling_factors") or self.sin is None:
             self.prepare_cos_sin(positions, offsets)
         num_tokens = positions.shape[0] * positions.shape[1]
         # HPU RoPE kernel requires hidden dimension for cos and sin to be equal