[feat]Ascend NPU Gemma-3-12b and Gemma-3-27b support (#8909)

python/sglang/srt/layers/activation.py

@@ -103,6 +103,15 @@ class GeluAndMul(CustomOp):
             raise RuntimeError("GeluAndMul only support tanh or none")
         return out
 
+    def forward_npu(self, x: torch.Tensor) -> torch.Tensor:
+        y_npu, gelu_npu = torch_npu.npu_geglu(
+            x,
+            dim=-1,
+            approximate=1 if self.approximate == "tanh" else 0,
+            activate_left=True,
+        )
+        return y_npu
+
 
 class NewGELU(CustomOp):
     def forward_native(self, x: torch.Tensor) -> torch.Tensor:
@@ -137,6 +146,9 @@ class QuickGELU(CustomOp):
         gelu_quick(x, out)
         return out
 
+    def forward_npu(self, x: torch.Tensor) -> torch.Tensor:
+        return torch_npu.npu_fast_gelu(x)
+
 
 class ScaledActivation(nn.Module):
     """An activation function with post-scale parameters.

python/sglang/srt/layers/attention/ascend_backend.py

@@ -180,7 +180,7 @@ class AscendAttnBackend(AttentionBackend):
 
             if self.use_fia:
                 """FIA will support multi-bs in the later version of CANN"""
-                q = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+                q = q.reshape(-1, layer.tp_q_head_num, layer.qk_head_dim)
                 attn_output = torch.empty(
                     (q.size(0), layer.tp_q_head_num, layer.v_head_dim),
                     device=q.device,
@@ -208,7 +208,8 @@ class AscendAttnBackend(AttentionBackend):
                 )
 
             else:
-                query = q.view(-1, layer.tp_q_head_num * layer.qk_head_dim)
+                if layer.qk_head_dim <= 128:
+                    query = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
                     attn_output = torch.empty(
                         (query.shape[0], layer.tp_q_head_num * layer.v_head_dim),
                         dtype=query.dtype,
@@ -228,6 +229,40 @@ class AscendAttnBackend(AttentionBackend):
                         num_kv_heads=layer.tp_k_head_num,
                         out=attn_output,
                     )
+                else:
+                    if layer.qk_head_dim != layer.v_head_dim:
+                        attn_output = q.new_empty(
+                            (q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
+                        )
+                    else:
+                        attn_output = torch.empty_like(q)
+
+                    use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
+
+                    q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+                    o_ = attn_output.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+
+                    causal = True
+                    if (
+                        layer.is_cross_attention
+                        or layer.attn_type == AttentionType.ENCODER_ONLY
+                    ):
+                        causal = False
+
+                    self.native_attn._run_sdpa_forward_extend(
+                        q_,
+                        o_,
+                        k_cache.view(-1, layer.tp_k_head_num, layer.qk_head_dim),
+                        v_cache.view(-1, layer.tp_v_head_num, layer.v_head_dim),
+                        forward_batch.req_to_token_pool.req_to_token,
+                        forward_batch.req_pool_indices,
+                        forward_batch.seq_lens,
+                        forward_batch.extend_prefix_lens,
+                        forward_batch.extend_seq_lens,
+                        scaling=layer.scaling,
+                        enable_gqa=use_gqa,
+                        causal=causal,
+                    )
         else:
             assert (
                 layer.qk_head_dim != layer.v_head_dim
@@ -283,7 +318,7 @@ class AscendAttnBackend(AttentionBackend):
             v_cache = forward_batch.token_to_kv_pool.get_value_buffer(
                 layer.layer_id
             ).view(-1, self.page_size, layer.tp_v_head_num * layer.v_head_dim)
-            query = q.view(-1, 1, layer.tp_q_head_num * layer.qk_head_dim)
+            query = q.reshape(-1, 1, layer.tp_q_head_num * layer.qk_head_dim)
             if self.forward_metadata.seq_lens_cpu_int is None:
                 actual_seq_len_kv = self.forward_metadata.seq_lens_cpu_list
             else:
@@ -439,7 +474,8 @@ class AscendAttnBackend(AttentionBackend):
                     scale=layer.scaling,
                 )
             else:
-                query = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+                query = q.reshape(-1, layer.tp_q_head_num, layer.qk_head_dim)
+                num_tokens = query.shape[0]
                 attn_output = torch.empty(
                     (num_tokens, layer.tp_q_head_num, layer.v_head_dim),
                     dtype=query.dtype,

python/sglang/srt/layers/layernorm.py

@@ -53,7 +53,7 @@ elif _is_hip:
 
 logger = logging.getLogger(__name__)
 
-if is_npu():
+if _is_npu:
     import torch_npu
 
 
@@ -266,23 +266,48 @@ class GemmaRMSNorm(CustomOp):
         out = gemma_rmsnorm(x, self.weight.data, self.variance_epsilon)
         return out
 
+    def forward_npu(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        orig_dtype = x.dtype
+        if residual is not None:
+            x = x + residual
+            residual = x
 
-class Gemma3RMSNorm(nn.Module):
+        x = x.float()
+        variance = torch_npu.mean(torch_npu.pow(x, 2), dim=-1, keepdim=True)
+        x = x * torch_npu.rsqrt(variance + self.variance_epsilon)
+        x = x * (1.0 + self.weight.float())
+        x = x.to(orig_dtype)
+        return x if residual is None else (x, residual)
+
+
+class Gemma3RMSNorm(CustomOp):
     def __init__(self, dim: int, eps: float = 1e-6):
         super().__init__()
         self.eps = eps
         self.weight = nn.Parameter(torch.zeros(dim))
+        # Re-dispatch
 
     def _norm(self, x):
         return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
 
-    def forward(self, x):
+    def forward_native(self, x):
         output = self._norm(x.float())
         # Llama does x.to(float16) * w whilst Gemma3 is (x * w).to(float16)
         # See https://github.com/huggingface/transformers/pull/29402
         output = output * (1.0 + self.weight.float())
         return output.type_as(x)
 
+    def forward_cuda(self, x):
+        return self.forward_native(x)
+
+    def forward_npu(self, x):
+        output, _ = torch_npu.npu_gemma_rms_norm(x, self.weight, self.eps)
+        return output
+
     def extra_repr(self):
         return f"{tuple(self.weight.shape)}, eps={self.eps}"
 

python/sglang/srt/layers/rotary_embedding.py

@@ -1876,7 +1876,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def apply_rotary_pos_emb(
+def apply_rotary_pos_emb_native(
     q: torch.Tensor,
     k: torch.Tensor,
     cos: torch.Tensor,
@@ -1899,6 +1899,33 @@ def apply_rotary_pos_emb(
     return q_embed, k_embed
 
 
+def apply_rotary_pos_emb_npu(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    unsqueeze_dim=1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if q.shape[1] != 128:
+        return apply_rotary_pos_emb_native(q, k, cos, sin, unsqueeze_dim)
+    cos = cos.unsqueeze(unsqueeze_dim)
+    cos = torch.transpose(cos, 1, 2)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    sin = torch.transpose(sin, 1, 2)
+    q = torch.transpose(q, 1, 2)
+    k = torch.transpose(k, 1, 2)
+    q_embed, k_embed = torch_npu.npu_apply_rotary_pos_emb(q, k, cos, sin)
+    q_embed = torch.transpose(q_embed, 1, 2)
+    k_embed = torch.transpose(k_embed, 1, 2)
+    return q_embed, k_embed
+
+
+if _is_npu:
+    apply_rotary_pos_emb = apply_rotary_pos_emb_npu
+else:
+    apply_rotary_pos_emb = apply_rotary_pos_emb_native
+
+
 def get_rope_cpu(
     head_size: int,
     rotary_dim: int,

python/sglang/srt/managers/mm_utils.py

@@ -20,9 +20,11 @@ from sglang.srt.managers.schedule_batch import (
 )
 from sglang.srt.mem_cache.multimodal_cache import MultiModalCache
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.utils import flatten_nested_list, print_warning_once
+from sglang.srt.utils import flatten_nested_list, is_npu, print_warning_once
 from sglang.utils import logger
 
+_is_npu = is_npu()
+
 # NOTE: Using the shared logger from sglang.utils instead of creating a module-specific logger
 # to ensure consistent logging behavior across the codebase. This prevents issues with log
 # propagation that can cause some log messages (like 'server is fired up') to not appear
@@ -486,6 +488,8 @@ def get_embedding_and_mask(
         if embedding is None:
             return None, None
     # 2. Get mask
+    if _is_npu:
+        torch.npu.current_stream().synchronize()
     special_multimodal_mask = _get_multimodal_mask(input_ids, placeholder_tensor)
     # 3. Adjust embedding length if needed
     embedding = _adjust_embedding_length(embedding, special_multimodal_mask, logger)

python/sglang/srt/multimodal/processors/base_processor.py

@@ -13,7 +13,9 @@ from PIL import Image
 from transformers import BaseImageProcessorFast
 
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-from sglang.srt.utils import load_audio, load_image, load_video, logger
+from sglang.srt.utils import is_npu, load_audio, load_image, load_video, logger
+
+_is_npu = is_npu()
 
 
 @dataclasses.dataclass
@@ -232,7 +234,7 @@ class BaseMultimodalProcessor(ABC):
             and isinstance(processor.image_processor, BaseImageProcessorFast)
             and not self.server_args.disable_fast_image_processor
         ):
-            kwargs["device"] = "cuda"
+            kwargs["device"] = "cuda" if not _is_npu else "npu"
         result = processor.__call__(
             text=[input_text],
             padding=True,