Merge pull request #6466 from vladmandic/api-get-memory

Implement API get-memory

modules/api/api.py

@@ -135,6 +135,7 @@ class Api:
         self.add_api_route("/sdapi/v1/preprocess", self.preprocess, methods=["POST"], response_model=PreprocessResponse)
         self.add_api_route("/sdapi/v1/train/embedding", self.train_embedding, methods=["POST"], response_model=TrainResponse)
         self.add_api_route("/sdapi/v1/train/hypernetwork", self.train_hypernetwork, methods=["POST"], response_model=TrainResponse)
+        self.add_api_route("/sdapi/v1/memory", self.get_memory, methods=["GET"], response_model=MemoryResponse)
 
     def add_api_route(self, path: str, endpoint, **kwargs):
         if shared.cmd_opts.api_auth:
@@ -501,6 +502,40 @@ class Api:
             shared.state.end()
             return TrainResponse(info = "train embedding error: {error}".format(error = error))
 
+    def get_memory(self):
+        try:
+            import os, psutil
+            process = psutil.Process(os.getpid())
+            res = process.memory_info() # only rss is cross-platform guaranteed so we dont rely on other values
+            ram_total = 100 * res.rss / process.memory_percent() # and total memory is calculated as actual value is not cross-platform safe
+            ram = { 'free': ram_total - res.rss, 'used': res.rss, 'total': ram_total }
+        except Exception as err:
+            ram = { 'error': f'{err}' }
+        try:
+            import torch
+            if torch.cuda.is_available():
+                s = torch.cuda.mem_get_info()
+                system = { 'free': s[0], 'used': s[1] - s[0], 'total': s[1] }
+                s = dict(torch.cuda.memory_stats(shared.device))
+                allocated = { 'current': s['allocated_bytes.all.current'], 'peak': s['allocated_bytes.all.peak'] }
+                reserved = { 'current': s['reserved_bytes.all.current'], 'peak': s['reserved_bytes.all.peak'] }
+                active = { 'current': s['active_bytes.all.current'], 'peak': s['active_bytes.all.peak'] }
+                inactive = { 'current': s['inactive_split_bytes.all.current'], 'peak': s['inactive_split_bytes.all.peak'] }
+                warnings = { 'retries': s['num_alloc_retries'], 'oom': s['num_ooms'] }
+                cuda = {
+                    'system': system,
+                    'active': active,
+                    'allocated': allocated,
+                    'reserved': reserved,
+                    'inactive': inactive,
+                    'events': warnings,
+                }
+            else:
+                cuda = { 'error': 'unavailable' }
+        except Exception as err:
+            cuda = { 'error': f'{err}' }
+        return MemoryResponse(ram = ram, cuda = cuda)
+
     def launch(self, server_name, port):
         self.app.include_router(self.router)
         uvicorn.run(self.app, host=server_name, port=port)

modules/api/models.py

@@ -260,3 +260,7 @@ class EmbeddingItem(BaseModel):
 class EmbeddingsResponse(BaseModel):
     loaded: Dict[str, EmbeddingItem] = Field(title="Loaded", description="Embeddings loaded for the current model")
     skipped: Dict[str, EmbeddingItem] = Field(title="Skipped", description="Embeddings skipped for the current model (likely due to architecture incompatibility)")
+
+class MemoryResponse(BaseModel):
+    ram: dict[str, str] | dict[str, float] = Field(title="RAM", description="System memory stats")
+    cuda: dict[str, str] | dict[str, dict] = Field(title="CUDA", description="nVidia CUDA memory stats")