matrice-compute 0.1.23__py3-none-any.whl → 0.1.24__py3-none-any.whl

matrice_compute/action_instance.py

@@ -285,13 +285,13 @@ class ActionInstance:
         ).get("gpuMemory", 0)
 
         logging.info(
-            "Action %s requires GPU with %d MB memory - selecting best-fit GPU(s)",
+            "Action %s requires GPU with %d MB memory - selecting GPU(s) with most free memory",
             action_id,
             required_memory
         )
 
         try:
-            # Get the best-fit GPU(s) with sufficient memory
+            # Get the GPU(s) with most free memory that have sufficient memory
             gpu_indices = get_gpu_with_sufficient_memory_for_action(
                 action_details=action_details
             )
@@ -1463,7 +1463,7 @@ def model_deploy_execute(self: ActionInstance):
     )
 
     # Get GPU configuration based on requirements and availability
-    # This uses the best-fit algorithm to select the most appropriate GPU(s)
+    # This selects the GPU(s) with the most free memory to balance load
     use_gpu = self.get_gpu_config(action_details)
 
     logging.info(

matrice_compute/instance_utils.py

@@ -600,13 +600,18 @@ def is_allowed_gpu_device(gpu_index: int) -> bool:
     Returns:
         bool: True if GPU is allowed (or no filter is set), False otherwise
     """
-    gpus = os.environ.get("GPUS")
-    if not gpus:
-        # No filter set - all GPUs are allowed
+    gpus = os.environ.get("GPUS", "").strip()
+    # No filter set or empty string - all GPUs are allowed
+    if not gpus or gpus == '""' or gpus == "''":
         return True
 
     try:
-        allowed_gpus = [int(x) for x in gpus.split(",") if x.strip()]
+        allowed_gpus = [int(x.strip()) for x in gpus.split(",") if x.strip()]
+
+        # If no valid GPUs after parsing, allow all
+        if not allowed_gpus:
+            return True
+
         is_allowed = int(gpu_index) in allowed_gpus
 
         if not is_allowed:
@@ -727,14 +732,14 @@ def get_gpu_with_sufficient_memory_for_action(
     # For smaller memory requirements, try to fit on a single GPU first
     if required_gpu_memory < 80000:
         logging.debug(
-            "Action %s: Required memory %d MB < 80000 MB - attempting single GPU allocation",
+            "Action %s: Required memory %d MB < 80000 MB - attempting single GPU allocation (selecting GPU with most free memory)",
             action_id,
             required_gpu_memory
         )
         try:
             single_gpu = get_single_gpu_with_sufficient_memory_for_action(action_details)
             logging.info(
-                "Action %s: Successfully allocated single GPU: %s",
+                "Action %s: Successfully allocated single GPU with most free memory: %s",
                 action_id,
                 single_gpu
             )
@@ -800,10 +805,10 @@ def get_single_gpu_with_sufficient_memory_for_action(
     action_details: dict,
 ) -> list:
     """
-    Get single GPU with sufficient memory using best-fit algorithm.
+    Get single GPU with sufficient memory using most-free algorithm.
 
-    Best-fit selects the GPU with the smallest amount of free memory
-    that still meets the requirements, minimizing fragmentation.
+    Selects the GPU with the MOST free memory that meets the requirements,
+    to balance load across GPUs and prevent any single GPU from being overused.
 
     Args:
         action_details (dict): Action details
@@ -818,7 +823,7 @@ def get_single_gpu_with_sufficient_memory_for_action(
     required_gpu_memory = get_required_gpu_memory(action_details)
 
     logging.debug(
-        "Action %s: Finding best-fit single GPU for %d MB",
+        "Action %s: Finding GPU with most free memory for %d MB",
         action_id,
         required_gpu_memory
     )
@@ -862,9 +867,9 @@ def get_single_gpu_with_sufficient_memory_for_action(
     if not memory_free_values:
         raise ValueError("No GPU devices found")
 
-    # Best-fit algorithm: find GPU with minimum free memory that meets requirement
+    # Most-free algorithm: find GPU with MAXIMUM free memory that meets requirement
     best_fit_gpu = None
-    best_fit_memory = float("inf")
+    best_fit_memory = 0  # Changed from float("inf") to 0
 
     for i, mem in enumerate(memory_free_values):
         # Check if GPU is in allowed list
@@ -887,12 +892,12 @@ def get_single_gpu_with_sufficient_memory_for_action(
                 required_gpu_memory
             )
 
-            # Best-fit: choose GPU with smallest sufficient memory
-            if mem < best_fit_memory:
+            # Most-free: choose GPU with MOST free memory to balance load
+            if mem > best_fit_memory:  # Changed from < to >
                 best_fit_gpu = i
                 best_fit_memory = mem
                 logging.debug(
-                    "Action %s: GPU %d is new best-fit candidate",
+                    "Action %s: GPU %d is new best candidate (most free memory)",
                     action_id,
                     i
                 )
@@ -907,7 +912,7 @@ def get_single_gpu_with_sufficient_memory_for_action(
 
     if best_fit_gpu is not None:
         logging.info(
-            "Action %s: Selected best-fit GPU %d with %d MB free (required: %d MB, waste: %d MB)",
+            "Action %s: Selected GPU %d with most free memory: %d MB free (required: %d MB, available: %d MB)",
             action_id,
             best_fit_gpu,
             best_fit_memory,

{matrice_compute-0.1.23.dist-info → matrice_compute-0.1.24.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: matrice_compute
-Version: 0.1.23
+Version: 0.1.24
 Summary: Common server utilities for Matrice.ai services
 Author-email: "Matrice.ai" <dipendra@matrice.ai>
 License-Expression: MIT

{matrice_compute-0.1.23.dist-info → matrice_compute-0.1.24.dist-info}/RECORD

@@ -1,17 +1,17 @@
 matrice_compute/__init__.py,sha256=ZzQcFsT005VCgq9VZUh565f4upOooEb_FwZ6RgweNZs,597
-matrice_compute/action_instance.py,sha256=kByPDNzmq93RBhVdnhTqGRLj7JleKFnH9hnIoJo966o,66215
+matrice_compute/action_instance.py,sha256=NK_ZWvNDrLUeOzWwXjxrX7XP-lDHbx5-A0K8ByFpnUg,66241
 matrice_compute/actions_manager.py,sha256=5U-xM6tl_Z6x96bi-c7AJM9ru80LqTN8f5Oce8dAu_A,7780
 matrice_compute/actions_scaledown_manager.py,sha256=pJ0nduNwHWZ10GnqJNx0Ok7cVWabQ_M8E2Vb9pH3A_k,2002
 matrice_compute/instance_manager.py,sha256=8USyX09ZxLvnVNIrjRogbyUeMCfgWnasuRqYkkVF4tQ,10146
-matrice_compute/instance_utils.py,sha256=tCI_A3L5iohw62acmlXuOJns0DjIkvwN4znlUAIkfbg,37863
+matrice_compute/instance_utils.py,sha256=xDOLo21G7unvlGTpnYQkEWSkyuAsVAcs4scOHy5Oxi4,38204
 matrice_compute/prechecks.py,sha256=W9YmNF3RcLhOf4U8WBlExvFqDw1aGWSNTlJtA73lbDQ,17196
 matrice_compute/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 matrice_compute/resources_tracker.py,sha256=pkdt0aVKx_TpY_Sq---73w9INkDffZZe3mZGlp1EftE,22573
 matrice_compute/scaling.py,sha256=CeT_lxJNkjJamRETG1lWaOtdSr5ySmcaMcqt7-lFRbo,23731
 matrice_compute/shutdown_manager.py,sha256=0MYV_AqygqR9NEntYf7atUC-PbWXyNkm1f-8c2aizgA,13234
 matrice_compute/task_utils.py,sha256=3qIutiQdYPyGRxH9ZwLbqdg8sZcnp6jp08pszWCRFl0,2820
-matrice_compute-0.1.23.dist-info/licenses/LICENSE.txt,sha256=_uQUZpgO0mRYL5-fPoEvLSbNnLPv6OmbeEDCHXhK6Qc,1066
-matrice_compute-0.1.23.dist-info/METADATA,sha256=7FCjLIs4y-5IfN9P8FRdcSbIZhPbeOC8Cg9ZSCUWr6o,1038
-matrice_compute-0.1.23.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-matrice_compute-0.1.23.dist-info/top_level.txt,sha256=63Plr3L1GzBUWZO5JZaFkiv8IcB10xUPU-9w3i6ptvE,16
-matrice_compute-0.1.23.dist-info/RECORD,,
+matrice_compute-0.1.24.dist-info/licenses/LICENSE.txt,sha256=_uQUZpgO0mRYL5-fPoEvLSbNnLPv6OmbeEDCHXhK6Qc,1066
+matrice_compute-0.1.24.dist-info/METADATA,sha256=5fsmPC37r0KPPd6h0qQXnvm0dFqLqboVInQdv7KCr5Y,1038
+matrice_compute-0.1.24.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+matrice_compute-0.1.24.dist-info/top_level.txt,sha256=63Plr3L1GzBUWZO5JZaFkiv8IcB10xUPU-9w3i6ptvE,16
+matrice_compute-0.1.24.dist-info/RECORD,,