Merge edd64d88f5 into 1daa72fa40

generation_inference.py

@@ -21,6 +21,9 @@ import torch
 from transformers import AutoModelForCausalLM
 
 from janus.models import MultiModalityCausalLM, VLChatProcessor
+from janus.utils.cuda_memory_manager import (
+    monitor_memory,
+)
 import numpy as np
 import os
 import PIL.Image
@@ -51,6 +54,7 @@ sft_format = vl_chat_processor.apply_sft_template_for_multi_turn_prompts(
 prompt = sft_format + vl_chat_processor.image_start_tag
 
 
+@monitor_memory(warning_threshold_gb=1.5, track_stats=True)
 @torch.inference_mode()
 def generate(
     mmgpt: MultiModalityCausalLM,
@@ -74,10 +78,16 @@ def generate(
 
     inputs_embeds = mmgpt.language_model.get_input_embeddings()(tokens)
 
-    generated_tokens = torch.zeros((parallel_size, image_token_num_per_image), dtype=torch.int).cuda()
+    generated_tokens = torch.zeros(
+        (parallel_size, image_token_num_per_image), dtype=torch.int
+    ).cuda()
 
     for i in range(image_token_num_per_image):
-        outputs = mmgpt.language_model.model(inputs_embeds=inputs_embeds, use_cache=True, past_key_values=outputs.past_key_values if i != 0 else None)
+        outputs = mmgpt.language_model.model(
+            inputs_embeds=inputs_embeds,
+            use_cache=True,
+            past_key_values=outputs.past_key_values if i != 0 else None,
+        )
         hidden_states = outputs.last_hidden_state
 
         logits = mmgpt.gen_head(hidden_states[:, -1, :])
@@ -90,12 +100,16 @@ def generate(
         next_token = torch.multinomial(probs, num_samples=1)
         generated_tokens[:, i] = next_token.squeeze(dim=-1)
 
-        next_token = torch.cat([next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1).view(-1)
+        next_token = torch.cat(
+            [next_token.unsqueeze(dim=1), next_token.unsqueeze(dim=1)], dim=1
+        ).view(-1)
         img_embeds = mmgpt.prepare_gen_img_embeds(next_token)
         inputs_embeds = img_embeds.unsqueeze(dim=1)
 
-
-    dec = mmgpt.gen_vision_model.decode_code(generated_tokens.to(dtype=torch.int), shape=[parallel_size, 8, img_size//patch_size, img_size//patch_size])
+    dec = mmgpt.gen_vision_model.decode_code(
+        generated_tokens.to(dtype=torch.int),
+        shape=[parallel_size, 8, img_size // patch_size, img_size // patch_size],
+    )
     dec = dec.to(torch.float32).cpu().numpy().transpose(0, 2, 3, 1)
 
     dec = np.clip((dec + 1) / 2 * 255, 0, 255)
@@ -103,9 +117,9 @@ def generate(
     visual_img = np.zeros((parallel_size, img_size, img_size, 3), dtype=np.uint8)
     visual_img[:, :, :] = dec
 
-    os.makedirs('generated_samples', exist_ok=True)
+    os.makedirs("generated_samples", exist_ok=True)
     for i in range(parallel_size):
-        save_path = os.path.join('generated_samples', "img_{}.jpg".format(i))
+        save_path = os.path.join("generated_samples", "img_{}.jpg".format(i))
         PIL.Image.fromarray(visual_img[i]).save(save_path)
 
 

janus/janusflow/models/clip_encoder.py

@@ -25,6 +25,9 @@ import torchvision.transforms
 from einops import rearrange
 
 from janus.janusflow.models.siglip_vit import create_siglip_vit
+from janus.utils.cuda_memory_manager import (
+    monitor_memory,
+)
 
 
 class CLIPVisionTower(nn.Module):
@@ -104,6 +107,7 @@ class CLIPVisionTower(nn.Module):
             raise ValueError(f"Unexpected select feature: {self.select_feature}")
         return image_features
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def forward(self, images):
         """
 

janus/janusflow/models/modeling_vlm.py

@@ -31,6 +31,7 @@ from transformers import (
 from transformers.models.llama.modeling_llama import LlamaRMSNorm
 from janus.janusflow.models.clip_encoder import CLIPVisionTower
 from janus.janusflow.models.uvit import ShallowUViTEncoder, ShallowUViTDecoder
+from janus.utils.cuda_memory_manager import monitor_memory
 import torch.nn as nn
 
 
@@ -168,6 +169,7 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
         )
         self.vision_gen_dec_aligner = nn.Linear(2048, 768, bias=True)
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def prepare_inputs_embeds(
         self,
         input_ids: torch.LongTensor,

janus/janusflow/models/processing_vlm.py

@@ -27,6 +27,7 @@ from transformers.processing_utils import ProcessorMixin
 
 from janus.janusflow.models.image_processing_vlm import VLMImageProcessor
 from janus.utils.conversation import get_conv_template
+from janus.utils.cuda_memory_manager import monitor_memory
 
 
 class DictOutput(object):
@@ -384,6 +385,7 @@ class VLChatProcessor(ProcessorMixin):
 
         return prepare
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def batchify(
         self, prepare_list: List[VLChatProcessorOutput]
     ) -> BatchedVLChatProcessorOutput:

janus/models/clip_encoder.py

@@ -25,6 +25,9 @@ import torchvision.transforms
 from einops import rearrange
 
 from janus.models.siglip_vit import create_siglip_vit
+from janus.utils.cuda_memory_manager import (
+    monitor_memory,
+)
 
 
 class CLIPVisionTower(nn.Module):
@@ -104,6 +107,7 @@ class CLIPVisionTower(nn.Module):
             raise ValueError(f"Unexpected select feature: {self.select_feature}")
         return image_features
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def forward(self, images):
         """
 

janus/models/modeling_vlm.py

@@ -31,6 +31,9 @@ from transformers.configuration_utils import PretrainedConfig
 
 from janus.models.clip_encoder import CLIPVisionTower
 from janus.models.projector import MlpProjector
+from janus.utils.cuda_memory_manager import (
+    monitor_memory,
+)
 
 
 class vision_head(torch.nn.Module):
@@ -218,6 +221,7 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
         language_config = config.language_config
         self.language_model = LlamaForCausalLM(language_config)
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def prepare_inputs_embeds(
         self,
         input_ids: torch.LongTensor,
@@ -259,6 +263,7 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
 
         return inputs_embeds
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def prepare_gen_img_embeds(self, image_ids: torch.LongTensor):
         return self.gen_aligner(self.gen_embed(image_ids))
 

janus/models/processing_vlm.py

@@ -27,6 +27,9 @@ from transformers.processing_utils import ProcessorMixin
 
 from janus.models.image_processing_vlm import VLMImageProcessor
 from janus.utils.conversation import get_conv_template
+from janus.utils.cuda_memory_manager import (
+    monitor_memory,
+)
 
 
 class DictOutput(object):
@@ -354,6 +357,7 @@ class VLChatProcessor(ProcessorMixin):
 
         return prepare
 
+    @monitor_memory(warning_threshold_gb=1.5, track_stats=True)
     def batchify(
         self, prepare_list: List[VLChatProcessorOutput]
     ) -> BatchedVLChatProcessorOutput:

janus/utils/cuda_memory_manager.py

@@ -0,0 +1,68 @@
+from functools import wraps
+from typing import Callable, Any
+import torch
+import warnings
+
+
+def monitor_memory(
+    warning_threshold_gb: float = 2.0,
+    track_stats: bool = True,
+    cleanup_on_warning: bool = True,
+) -> Callable:
+    """Memory monitoring decorator for CUDA operations.
+
+    Args:
+        warning_threshold_gb: Memory threshold in GB to trigger warnings
+        track_stats: Whether to track and print memory statistics
+        cleanup_on_warning: Whether to attempt memory cleanup when threshold is reached
+
+    Returns:
+        Decorator function that monitors memory usage
+    """
+
+    def decorator(func: Callable) -> Callable:
+        @wraps(func)
+        def wrapper(*args, **kwargs) -> Any:
+            if not torch.cuda.is_available():
+                return func(*args, **kwargs)
+
+            # Get initial memory state
+            free_before = torch.cuda.mem_get_info()[0] / 1024**3
+
+            try:
+                # Check memory state and cleanup if needed
+                if free_before < warning_threshold_gb and cleanup_on_warning:
+                    torch.cuda.empty_cache()
+                    free_after_cleanup = torch.cuda.mem_get_info()[0] / 1024**3
+
+                    if free_after_cleanup < warning_threshold_gb:
+                        warnings.warn(
+                            f"Low memory in {func.__name__}: {free_after_cleanup:.2f}GB free"
+                        )
+
+                result = func(*args, **kwargs)
+
+                # Track memory statistics if enabled
+                if track_stats:
+                    peak = torch.cuda.max_memory_allocated() / 1024**3
+                    free_after = torch.cuda.mem_get_info()[0] / 1024**3
+                    print(
+                        f"Memory stats for {func.__name__}:\n"
+                        f"Peak: {peak:.2f}GB | Delta: {free_before - free_after:.2f}GB"
+                    )
+                    torch.cuda.reset_peak_memory_stats()
+
+                return result
+
+            except RuntimeError as e:
+                if "out of memory" in str(e):
+                    free = torch.cuda.mem_get_info()[0] / 1024**3
+                    raise RuntimeError(
+                        f"OOM in {func.__name__} with {free:.2f}GB free. "
+                        "Consider reducing batch size or image resolution."
+                    ) from e
+                raise
+
+        return wrapper
+
+    return decorator