LoRA: Ascend: Update ascend LoRA backend to support new kernels (#15912)

python/sglang/srt/lora/backend/ascend_backend.py

@@ -29,22 +29,17 @@ class AscendLoRABackend(BaseLoRABackend):
         _, weight_out_dim, _ = weights.shape
 
         output_tensor = torch.zeros(
-            (total_seq_len, weight_out_dim), dtype=x.dtype, device=x.device
+            (total_seq_len, weight_out_dim), dtype=torch.float, device=x.device
         )
-        torch.ops.npu.sgmv_shrink(
+        torch.ops.npu.sgemmv_shrink(
             x,
             weights,
             self.batch_info.weight_indices,
             self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            self.batch_info.scalings,
             output_tensor,
-            1.0,
         )
-        scaling = (
-            self.batch_info.scalings.gather(0, self.batch_info.weight_indices)
-            .repeat_interleave(self.batch_info.seg_lens, output_size=total_seq_len)
-            .unsqueeze(-1)
-        )
-        output_tensor *= scaling
 
         return output_tensor
 
@@ -52,6 +47,7 @@ class AscendLoRABackend(BaseLoRABackend):
         self,
         x: torch.Tensor,
         weights: torch.Tensor,
+        output_offset: torch.Tensor,
         base_output: torch.Tensor = None,
         *args,
         **kwargs,
@@ -66,14 +62,14 @@ class AscendLoRABackend(BaseLoRABackend):
         else:
             output_tensor = base_output
 
-        torch.ops.npu.sgmv_expand(
+        torch.ops.npu.sgemmv_expand(
             x,
             weights,
             self.batch_info.weight_indices,
             self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            output_offset,
             output_tensor,
-            0,
-            weight_out_dim,
         )
 
         return output_tensor
@@ -84,8 +80,6 @@ class AscendLoRABackend(BaseLoRABackend):
         qkv_lora_a: torch.Tensor,
         qkv_lora_b: torch.Tensor,
         output_offset: torch.Tensor,
-        output_offset_cpu: torch.Tensor,
-        max_qkv_out_dim: int,
         base_output: torch.Tensor = None,
         n_slices: int = 3,
         *args,
@@ -106,37 +100,28 @@ class AscendLoRABackend(BaseLoRABackend):
             output_tensor = base_output
 
         lora_a_output = torch.zeros(
-            total_seq_len, weight_intermediate_dim, dtype=x.dtype, device=x.device
+            total_seq_len, weight_intermediate_dim, dtype=torch.float, device=x.device
         )
-        torch.ops.npu.sgmv_shrink(
+
+        torch.ops.npu.sgemmv_shrink(
             x,
             qkv_lora_a,
             self.batch_info.weight_indices,
             self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            self.batch_info.scalings,
             lora_a_output,
-            1.0,
         )
 
-        scaling = (
+        torch.ops.npu.sgemmv_expand(
-            self.batch_info.scalings.gather(0, self.batch_info.weight_indices)
+            lora_a_output,
-            .repeat_interleave(self.batch_info.seg_lens, output_size=total_seq_len)
+            qkv_lora_b,
-            .unsqueeze(-1)
+            self.batch_info.weight_indices,
+            self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            output_offset,
+            output_tensor,
         )
-        lora_a_output *= scaling
-
-        for slice_id in range(n_slices):
-            slice_offset = output_offset_cpu[slice_id]
-            slice_offset_next = output_offset_cpu[slice_id + 1]
-            slice_size = slice_offset_next - slice_offset
-            torch.ops.npu.sgmv_expand(
-                lora_a_output[:, (max_rank * slice_id) : (max_rank * (slice_id + 1))],
-                qkv_lora_b[:, slice_offset:slice_offset_next],
-                self.batch_info.weight_indices,
-                self.batch_info.seg_lens,
-                output_tensor,
-                slice_offset,
-                slice_size,
-            )
 
         return output_tensor
 
@@ -145,19 +130,17 @@ class AscendLoRABackend(BaseLoRABackend):
         x: torch.Tensor,
         gate_up_lora_a: torch.Tensor,
         gate_up_lora_b: torch.Tensor,
+        output_offset: torch.Tensor,
         base_output: torch.Tensor = None,
         *args,
         **kwargs,
     ) -> torch.Tensor:
 
-        num_slices = 2
         assert isinstance(gate_up_lora_b, torch.Tensor)
 
         total_seq_len, _ = x.shape
         _, weight_intermediate_dim, _ = gate_up_lora_a.shape
         _, weight_out_dim, _ = gate_up_lora_b.shape
-        slice_size = weight_out_dim // num_slices
-        max_rank = weight_intermediate_dim // num_slices
 
         if base_output is None:
             output_tensor = torch.zeros(
@@ -167,37 +150,28 @@ class AscendLoRABackend(BaseLoRABackend):
             output_tensor = base_output
 
         lora_a_output = torch.zeros(
-            total_seq_len, weight_intermediate_dim, dtype=x.dtype, device=x.device
+            total_seq_len, weight_intermediate_dim, dtype=torch.float, device=x.device
         )
 
-        torch.ops.npu.sgmv_shrink(
+        torch.ops.npu.sgemmv_shrink(
             x,
             gate_up_lora_a,
             self.batch_info.weight_indices,
             self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            self.batch_info.scalings,
             lora_a_output,
-            1.0,
         )
 
-        scaling = (
+        torch.ops.npu.sgemmv_expand(
-            self.batch_info.scalings.gather(0, self.batch_info.weight_indices)
+            lora_a_output,
-            .repeat_interleave(self.batch_info.seg_lens, output_size=total_seq_len)
+            gate_up_lora_b,
-            .unsqueeze(-1)
+            self.batch_info.weight_indices,
+            self.batch_info.seg_lens,
+            self.batch_info.lora_ranks,
+            output_offset,
+            output_tensor,
         )
-        lora_a_output *= scaling
-
-        slice_offset = 0
-        for slice_id in range(num_slices):
-            torch.ops.npu.sgmv_expand(
-                lora_a_output[:, (max_rank * slice_id) : (max_rank * (slice_id + 1))],
-                gate_up_lora_b[:, slice_offset : slice_offset + slice_size],
-                self.batch_info.weight_indices,
-                self.batch_info.seg_lens,
-                output_tensor,
-                slice_offset,
-                slice_size,
-            )
-            slice_offset += slice_size
 
         return output_tensor
 
@@ -218,7 +192,7 @@ class AscendLoRABackend(BaseLoRABackend):
                 max_len=num_tokens_per_req,
                 weight_indices=torch.zeros(max_bs_in_cuda_graph, dtype=torch.int32),
                 lora_ranks=torch.zeros(self.max_loras_per_batch, dtype=torch.int32),
-                scalings=torch.zeros(self.max_loras_per_batch, dtype=torch.float),
+                scalings=torch.zeros(self.max_loras_per_batch, dtype=torch.float16),
                 permutation=None,
             )
 
@@ -246,7 +220,7 @@ class AscendLoRABackend(BaseLoRABackend):
             lora_ranks, dtype=torch.int32, pin_memory=True, device="cpu"
         )
         scalings_tensor = torch.tensor(
-            scalings, dtype=torch.float, pin_memory=True, device="cpu"
+            scalings, dtype=torch.float16, pin_memory=True, device="cpu"
         )
 
         bs = forward_batch.batch_size
@@ -287,7 +261,7 @@ class AscendLoRABackend(BaseLoRABackend):
                     (self.max_loras_per_batch,), dtype=torch.int32, device=self.device
                 ),
                 scalings=torch.empty(
-                    (self.max_loras_per_batch,), dtype=torch.float, device=self.device
+                    (self.max_loras_per_batch,), dtype=torch.float16, device=self.device
                 ),
                 permutation=None,
             )