[Bugfix] Add fully sharded layer for QKVParallelLinearWithLora (#5665)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>

Author: jeejeelee

tests/lora/test_baichuan.py

@@ -64,7 +64,8 @@ def test_baichuan_lora(baichuan_lora_files):
 
 
 @pytest.mark.skip("Requires multiple GPUs")
-def test_baichuan_tensor_parallel_equality(baichuan_lora_files):
+@pytest.mark.parametrize("fully_sharded", [True, False])
+def test_baichuan_tensor_parallel_equality(baichuan_lora_files, fully_sharded):
     # Cannot use as it will initialize torch.cuda too early...
     # if torch.cuda.device_count() < 4:
     #     pytest.skip(f"Not enough GPUs for tensor parallelism {4}")
@@ -75,7 +76,8 @@ def test_baichuan_tensor_parallel_equality(baichuan_lora_files):
                        max_loras=4,
                        max_lora_rank=64,
                        tensor_parallel_size=1,
-                       trust_remote_code=True)
+                       trust_remote_code=True,
+                       fully_sharded_loras=fully_sharded)
     output_tp1 = do_sample(llm_tp1, baichuan_lora_files, lora_id=1)
 
     del llm_tp1
@@ -87,7 +89,8 @@ def test_baichuan_tensor_parallel_equality(baichuan_lora_files):
                        max_loras=4,
                        max_lora_rank=64,
                        tensor_parallel_size=2,
-                       trust_remote_code=True)
+                       trust_remote_code=True,
+                       fully_sharded_loras=fully_sharded)
     output_tp2 = do_sample(llm_tp2, baichuan_lora_files, lora_id=2)
 
     del llm_tp2
@@ -101,10 +104,11 @@ def test_baichuan_tensor_parallel_equality(baichuan_lora_files):
                        max_loras=4,
                        max_lora_rank=64,
                        tensor_parallel_size=4,
-                       trust_remote_code=True)
+                       trust_remote_code=True,
+                       fully_sharded_loras=fully_sharded)
     output_tp4 = do_sample(llm_tp4, baichuan_lora_files, lora_id=2)
 
     del llm_tp4
     cleanup()
 
-    assert output_tp1 == output_tp4
+    assert output_tp1 == output_tp4

tests/lora/test_layers.py

@@ -12,7 +12,8 @@
 from vllm.lora.fully_sharded_layers import (
     ColumnParallelLinearWithShardedLoRA,
     MergedColumnParallelLinearWithShardedLoRA,
-    MergedQKVParallelLinearWithShardedLora, RowParallelLinearWithShardedLoRA)
+    MergedQKVParallelLinearWithShardedLora, QKVParallelLinearWithShardedLora,
+    RowParallelLinearWithShardedLoRA)
 # yapf conflicts with isort for this block
 # yapf: disable
 from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
@@ -684,7 +685,9 @@ def create_column_parallel_packed_layer():
                                        bias=False,
                                        params_dtype=torch.float16)
             linear.weight.data = torch.rand_like(linear.weight.data)
-            lora_linear = QKVParallelLinearWithLora(linear)
+            lora_linear = QKVParallelLinearWithLora(
+                linear
+            ) if not fully_shard else QKVParallelLinearWithShardedLora(linear)
 
         @dataclass
         class FakeConfig:

vllm/lora/fully_sharded_layers.py

@@ -12,6 +12,7 @@
 from vllm.lora.layers import (ColumnParallelLinearWithLoRA,
                               MergedColumnParallelLinearWithLoRA,
                               MergedQKVParallelLinearWithLora,
+                              QKVParallelLinearWithLora,
                               RowParallelLinearWithLoRA)
 from vllm.lora.punica import bgmv, dispatch_bgmv_low_level
 
@@ -90,11 +91,11 @@ def can_replace_layer(cls, source_layer: nn.Module,
 def _mcp_apply(x, bias, layer):
     """
     MergedColumnParallelLinearWithShardedLoRA and 
-    QKVParallelLinearWithShardedLora share the same 
+    MergedQKVParallelLinearWithShardedLora share the same 
     LoRa weight application method.
     
     The main difference is the step by shard_size for lora_b which can
-    vary for QKVParallelLinearWithShardedLora but is constant for 
+    vary for MergedQKVParallelLinearWithShardedLora but is constant for 
     MergedColumnParallelLinearWithShardedLoRA.
     """
     # expecting 2 for column parallel and 3 for qkv
@@ -167,14 +168,65 @@ def can_replace_layer(cls, source_layer: nn.Module,
         )
 
 
-class MergedQKVParallelLinearWithShardedLora(MergedQKVParallelLinearWithLora):
+class QKVParallelLinearWithShardedLora(QKVParallelLinearWithLora):
     """
     Differs from QKVParallelLinearWithLora by slicing the 
     LoRA A's also.
 
     Based on S-LoRA, slicing happens along the rank dim.
     """
 
+    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
+        tp_rank = get_tensor_model_parallel_rank()
+        shard_size = self.lora_a_stacked.shape[2]
+        start_idx = tp_rank * shard_size
+        lora_a = lora_a[:, start_idx:start_idx + shard_size]
+        return lora_a
+
+    def apply(self, x: torch.Tensor,
+              bias: Optional[torch.Tensor]) -> torch.Tensor:
+        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
+
+        x = x.view(-1, x.shape[-1])
+        output, out_orig_shape = output.view(-1,
+                                             output.shape[-1]), output.shape
+        buffer = torch.zeros((x.shape[0], self.lora_a_stacked.shape[2]),
+                             dtype=torch.float32,
+                             device=x.device)
+
+        bgmv(buffer, x, self.lora_a_stacked,
+             self.indices[:self.indices_len[0]], 0, 1.0)
+        buffer = tensor_model_parallel_all_gather(buffer)
+        bgmv(output, buffer, self.lora_b_stacked,
+             self.indices[:self.indices_len[0]], 0, 1.0)
+        # now have column partitioned output
+
+        output = output.view(*out_orig_shape)
+        return output
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(cls, source_layer: nn.Module,
+                          lora_config: LoRAConfig, packed_modules_list: List,
+                          model_config: Optional[PretrainedConfig]) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )
+
+
+class MergedQKVParallelLinearWithShardedLora(MergedQKVParallelLinearWithLora):
+    """
+    Differs from MergedQKVParallelLinearWithLora by slicing the 
+    LoRA A's also.
+
+    Based on S-LoRA, slicing happens along the rank dim.
+    """
+
     def slice_lora_a(
         self, lora_a: List[Union[torch.Tensor, None]]
     ) -> List[Union[torch.Tensor, None]]:

vllm/lora/layers.py

@@ -641,6 +641,24 @@ def __init__(self, base_layer: QKVParallelLinear) -> None:
         self.kv_proj_total_size = (self.base_layer.total_num_kv_heads *
                                    self.base_layer.head_size)
 
+    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
+        tp_rank = get_tensor_model_parallel_rank()
+        self.q_shard_id = tp_rank
+        self.kv_shard_id = tp_rank // self.base_layer.num_kv_head_replicas
+        lora_b_q = lora_b[:, self.q_proj_shard_size *
+                          self.q_shard_id:self.q_proj_shard_size *
+                          (self.q_shard_id + 1)]
+        k_offset = self.q_proj_total_size
+        lora_b_k = lora_b[:, k_offset +
+                          self.kv_proj_shard_size * self.kv_shard_id:k_offset +
+                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        v_offset = k_offset + self.kv_proj_total_size
+        lora_b_v = lora_b[:, v_offset +
+                          self.kv_proj_shard_size * self.kv_shard_id:v_offset +
+                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        lora_b = torch.cat([lora_b_q, lora_b_k, lora_b_v], dim=1)
+        return lora_b
+
     def set_lora(
         self,
         index: int,
@@ -650,21 +668,8 @@ def set_lora(
     ):
         self.reset_lora(index)
         if self.tp_size > 1:
-            tp_rank = get_tensor_model_parallel_rank()
-            self.q_shard_id = tp_rank
-            self.kv_shard_id = tp_rank // self.base_layer.num_kv_head_replicas
-            lora_b_q = lora_b[:, self.q_proj_shard_size *
-                              self.q_shard_id:self.q_proj_shard_size *
-                              (self.q_shard_id + 1)]
-            k_offset = self.q_proj_total_size
-            lora_b_k = lora_b[:, k_offset + self.kv_proj_shard_size *
-                              self.kv_shard_id:k_offset +
-                              self.kv_proj_shard_size * (self.kv_shard_id + 1)]
-            v_offset = k_offset + self.kv_proj_total_size
-            lora_b_v = lora_b[:, v_offset + self.kv_proj_shard_size *
-                              self.kv_shard_id:v_offset +
-                              self.kv_proj_shard_size * (self.kv_shard_id + 1)]
-            lora_b = torch.cat([lora_b_q, lora_b_k, lora_b_v], dim=1)
+            lora_a = self.slice_lora_a(lora_a)
+            lora_b = self.slice_lora_b(lora_b)
 
         self.lora_a_stacked[index,
                             0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
@@ -674,6 +679,7 @@ def set_lora(
                                 lora_b.T, non_blocking=True)
 
     @classmethod
+    @_not_fully_sharded_can_replace
     def can_replace_layer(cls, source_layer: nn.Module,
                           lora_config: LoRAConfig, packed_modules_list: List,
                           model_config: Optional[PretrainedConfig]) -> bool:

vllm/lora/utils.py

@@ -8,7 +8,8 @@
 from vllm.lora.fully_sharded_layers import (
     ColumnParallelLinearWithShardedLoRA,
     MergedColumnParallelLinearWithShardedLoRA,
-    MergedQKVParallelLinearWithShardedLora, RowParallelLinearWithShardedLoRA)
+    MergedQKVParallelLinearWithShardedLora, QKVParallelLinearWithShardedLora,
+    RowParallelLinearWithShardedLoRA)
 # being imported for _all_lora_classes below
 # yapf conflicts with isort for this block
 # yapf: disable
@@ -35,6 +36,7 @@
     RowParallelLinearWithLoRA,
     LogitsProcessorWithLoRA,
     ColumnParallelLinearWithShardedLoRA,
+    QKVParallelLinearWithShardedLora,
     MergedColumnParallelLinearWithShardedLoRA,
     MergedQKVParallelLinearWithShardedLora,
     RowParallelLinearWithShardedLoRA,