From 8c4bee49a773733fd4967c5ec4d37b2079386236 Mon Sep 17 00:00:00 2001
From: ryan <ryan@tmfi.us>
Date: Thu, 27 Mar 2025 11:23:15 -0700
Subject: [PATCH] deps: lazy load optional dependencies gguf, torchvision

---
 python/sglang/srt/configs/deepseekvl2.py       | 11 ++++++++++-
 python/sglang/srt/model_loader/loader.py       | 12 +++++++++++-
 python/sglang/srt/model_loader/weight_utils.py |  5 ++++-
 3 files changed, 25 insertions(+), 3 deletions(-)

diff --git a/python/sglang/srt/configs/deepseekvl2.py b/python/sglang/srt/configs/deepseekvl2.py
index 2740742031be..a04549224206 100644
--- a/python/sglang/srt/configs/deepseekvl2.py
+++ b/python/sglang/srt/configs/deepseekvl2.py
@@ -4,7 +4,6 @@
 from typing import Dict, List, Optional, Tuple
 
 import torch
-import torchvision.transforms as T
 from PIL import Image, ImageOps
 from transformers import (
     AutoProcessor,
@@ -76,6 +75,16 @@ def __init__(
         self.std = std
         self.normalize = normalize
 
+        # only load torchvision.transforms when needed
+        try:
+            import torchvision.transforms as T
+
+            # FIXME: add version check for gguf
+        except ImportError as err:
+            raise ImportError(
+                "Please install torchvision via `pip install torchvision` to use Deepseek-VL2."
+            ) from err
+
         transform_pipelines = [T.ToTensor()]
 
         if normalize:
diff --git a/python/sglang/srt/model_loader/loader.py b/python/sglang/srt/model_loader/loader.py
index e01eb8ad5801..07e582cc5d74 100644
--- a/python/sglang/srt/model_loader/loader.py
+++ b/python/sglang/srt/model_loader/loader.py
@@ -14,7 +14,6 @@
 from contextlib import contextmanager
 from typing import Any, Dict, Generator, Iterable, List, Optional, Tuple, cast
 
-import gguf
 import huggingface_hub
 import numpy as np
 import torch
@@ -1155,6 +1154,17 @@ def _get_gguf_weights_map(self, model_config: ModelConfig):
         See "Standardized tensor names" in
         https://github.com/ggerganov/ggml/blob/master/docs/gguf.md for details.
         """
+
+        # only load the gguf module when needed
+        try:
+            import gguf
+
+            # FIXME: add version check for gguf
+        except ImportError as err:
+            raise ImportError(
+                "Please install gguf via `pip install gguf` to use gguf quantizer."
+            ) from err
+
         config = model_config.hf_config
         model_type = config.model_type
         # hack: ggufs have a different name than transformers
diff --git a/python/sglang/srt/model_loader/weight_utils.py b/python/sglang/srt/model_loader/weight_utils.py
index 92751b3c130d..0fb975e657ce 100644
--- a/python/sglang/srt/model_loader/weight_utils.py
+++ b/python/sglang/srt/model_loader/weight_utils.py
@@ -22,7 +22,6 @@
 )
 
 import filelock
-import gguf
 import huggingface_hub.constants
 import numpy as np
 import safetensors.torch
@@ -464,6 +463,8 @@ def pt_weights_iterator(
 def get_gguf_extra_tensor_names(
     gguf_file: str, gguf_to_hf_name_map: Dict[str, str]
 ) -> List[str]:
+    import gguf
+
     reader = gguf.GGUFReader(gguf_file)
     expected_gguf_keys = set(gguf_to_hf_name_map.keys())
     exact_gguf_keys = set([tensor.name for tensor in reader.tensors])
@@ -479,6 +480,8 @@ def gguf_quant_weights_iterator(
     them to torch tensors
     """
 
+    import gguf
+
     reader = gguf.GGUFReader(gguf_file)
 
     for tensor in reader.tensors: