Fix model loader for more quantization formats (#2448)

python/sglang/srt/models/llama.py

@@ -294,6 +294,28 @@ class LlamaModel(nn.Module):
 
 
 class LlamaForCausalLM(nn.Module):
+
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    # in TP, these weights are partitioned along the column dimension (dim=-1)
+    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
     def __init__(
         self,
         config: LlamaConfig,

python/sglang/srt/models/qwen2.py

@@ -267,6 +267,26 @@ class Qwen2Model(nn.Module):
 
 
 class Qwen2ForCausalLM(nn.Module):
+
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
     def __init__(
         self,
         config: Qwen2Config,

python/sglang/srt/server_args.py

@@ -283,7 +283,15 @@ class ServerArgs:
             "--load-format",
             type=str,
             default=ServerArgs.load_format,
-            choices=["auto", "pt", "safetensors", "npcache", "dummy", "gguf"],
+            choices=[
+                "auto",
+                "pt",
+                "safetensors",
+                "npcache",
+                "dummy",
+                "gguf",
+                "bitsandbytes",
+            ],
             help="The format of the model weights to load. "
             '"auto" will try to load the weights in the safetensors format '
             "and fall back to the pytorch bin format if safetensors format "
@@ -294,7 +302,9 @@ class ServerArgs:
             "a numpy cache to speed up the loading. "
             '"dummy" will initialize the weights with random values, '
             "which is mainly for profiling."
-            '"gguf" will load the weights in the gguf format. ',
+            '"gguf" will load the weights in the gguf format. '
+            '"bitsandbytes" will load the weights using bitsandbytes '
+            "quantization.",
         )
         parser.add_argument(
             "--trust-remote-code",