[4/n] Support gpt oss 20b lora (#21570)

python/sglang/srt/lora/layers.py

@@ -629,7 +629,6 @@ class RowParallelLinearWithLoRA(BaseLayerWithLoRA):
         return lora_output
 
     def forward(self, input_: torch.Tensor, skip_all_reduce=False):
-        # duplicate the logic in RowParallelLinear
         if self.base_layer.input_is_parallel:
             input_parallel = input_
         else:
@@ -638,8 +637,14 @@ class RowParallelLinearWithLoRA(BaseLayerWithLoRA):
                 input_, num_partitions=self.base_layer.tp_size
             )
             input_parallel = splitted_input[tp_rank].contiguous()
+
+        bias_ = (
+            None
+            if (self.base_layer.tp_rank > 0 or self.base_layer.skip_bias_add)
+            else self.base_layer.bias
+        )
         output_parallel = self.base_layer.quant_method.apply(
-            self.base_layer, input_parallel
+            self.base_layer, input_parallel, bias=bias_
         )
 
         should_reduce = (
@@ -668,17 +673,8 @@ class RowParallelLinearWithLoRA(BaseLayerWithLoRA):
             else:
                 output_ = output_parallel
 
-        if not self.base_layer.skip_bias_add:
-            output = (
-                output_ + self.base_layer.bias
-                if self.base_layer.bias is not None
-                else output_
-            )
-            output_bias = None
-        else:
-            output = output_
-            output_bias = self.base_layer.bias
-        return output, output_bias
+        output_bias = self.base_layer.bias if self.base_layer.skip_bias_add else None
+        return output_, output_bias
 
     def slice_lora_a_weights(self, A: torch.Tensor, tp_rank: int):
         shard_size = self.base_layer.input_size_per_partition
@@ -719,6 +715,9 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
         self.intermediate_size_per_partition = getattr(
             base_layer, "intermediate_size_per_partition", None
         )
+        self._uses_interleaved_gate_up = (
+            getattr(base_layer.moe_runner_config, "gemm1_alpha", None) is not None
+        )
 
         # initialize triton_lora moe runner for batches with lora enabled
         from sglang.srt.layers.moe.moe_runner.runner import MoeRunner
@@ -895,7 +894,10 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
           gate_up_proj_moe B: [intermediate_size*2, rank] — output matches sharded base w13
           down_proj_moe B:    [hidden_size, rank] — output is all-reduced, no slice
         """
-        if self.tp_size <= 1:
+        needs_processing = (self.tp_size > 1) or (
+            target_module == "gate_up_proj_moe" and self._uses_interleaved_gate_up
+        )
+        if not needs_processing:
             return B
         if target_module != "gate_up_proj_moe":
             return B
@@ -923,6 +925,8 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
             full_inter = B.shape[0] // 2
             gate_b = B[start:end, :]
             up_b = B[full_inter + start : full_inter + end, :]
+            if self._uses_interleaved_gate_up:
+                return torch.stack([gate_b, up_b], dim=1).reshape(-1, B.shape[-1])
             return torch.cat([gate_b, up_b], dim=0).contiguous()
         return B
 

python/sglang/srt/lora/mem_pool.py

@@ -315,7 +315,7 @@ class LoRAMemoryPool:
                     # MoE expert version (4D)
                     moe_key = f"{module_name}_moe"
                     buffer[moe_key] = [
-                        torch.empty(
+                        torch.zeros(
                             get_lora_shape_fn(
                                 moe_key, base_model, self.max_lora_rank, idx
                             ),
@@ -327,7 +327,7 @@ class LoRAMemoryPool:
                 else:
                     # Standard allocation for unambiguous modules
                     buffer[module_name] = [
-                        torch.empty(
+                        torch.zeros(
                             get_lora_shape_fn(
                                 module_name,
                                 base_model,
@@ -347,7 +347,7 @@ class LoRAMemoryPool:
         ):
             target_modules = target_modules & set(EMBEDDING_NAMES)
             for module_name in target_modules:
-                buffer[module_name] = torch.empty(
+                buffer[module_name] = torch.zeros(
                     get_lora_shape_fn(
                         module_name,
                         base_model,
@@ -359,7 +359,7 @@ class LoRAMemoryPool:
                 )
 
         if self.lora_added_tokens_size > 0:
-            self.new_embeddings_buffer["input_embeddings"] = torch.empty(
+            self.new_embeddings_buffer["input_embeddings"] = torch.zeros(
                 (
                     self.max_loras_per_batch,
                     self.lora_added_tokens_size,

python/sglang/srt/lora/utils.py

@@ -88,9 +88,17 @@ def get_hidden_dim(
         elif module_name == "down_proj":
             return config.intermediate_size, config.hidden_size
         elif module_name == "gate_up_proj_moe":
-            return config.hidden_size, config.moe_intermediate_size * 2
+            moe_inter = (
+                getattr(config, "moe_intermediate_size", None)
+                or config.intermediate_size
+            )
+            return config.hidden_size, moe_inter * 2
         elif module_name == "down_proj_moe":
-            return config.moe_intermediate_size, config.hidden_size
+            moe_inter = (
+                getattr(config, "moe_intermediate_size", None)
+                or config.intermediate_size
+            )
+            return moe_inter, config.hidden_size
         elif module_name == "embed_tokens":
             # For embedding: input is vocab_size (as embedding lookup), output is hidden_size
             # if contain extra tokens will be added; otherwise is 0.

python/sglang/srt/models/gpt_oss.py

@@ -17,6 +17,7 @@
 
 import logging
 import math
+import re
 from collections.abc import Iterable
 from functools import partial
 from typing import Any, Dict, List, Optional, Tuple, Union
@@ -651,6 +652,13 @@ class GptOssModel(nn.Module):
 class GptOssForCausalLM(nn.Module):
     fall_back_to_pt_during_load = False
 
+    _lora_pattern_moe = re.compile(
+        r"^(?:model\.layers\.\d+\.(?:self_attn\.(?:qkv_proj|o_proj)|mlp\.experts)|lm_head|model\.embed_tokens)$"
+    )
+
+    def should_apply_lora(self, module_name: str) -> bool:
+        return bool(self._lora_pattern_moe.match(module_name))
+
     def __init__(
         self,
         config: GptOssConfig,