[Core] Clarify memory variable naming in model runner (#20060)

python/sglang/srt/model_executor/model_runner.py

@@ -391,8 +391,8 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         # Initialize MooncakeTransferEngine
         self.init_shared_mooncake_transfer_engine()
 
-        # Get memory before model loading
-        min_per_gpu_memory = self.init_torch_distributed()
+        # Get available memory before model loading
+        pre_model_load_memory = self.init_torch_distributed()
 
         # Init forward stream for overlap schedule
         self.forward_stream = torch.get_device_module(self.device).Stream()
@@ -413,7 +413,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             deep_gemm_wrapper.update_deep_gemm_config(gpu_id, server_args)
 
         # Initialize the model runner
-        self.initialize(min_per_gpu_memory)
+        self.initialize(pre_model_load_memory)
         self.check_quantized_moe_compatibility()
 
         if self.is_multimodal:
@@ -447,7 +447,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                 port=self.dist_port,
             )
 
-    def initialize(self, min_per_gpu_memory: float):
+    def initialize(self, pre_model_load_memory: float):
         server_args = self.server_args
 
         self.memory_saver_adapter = TorchMemorySaverAdapter.create(
@@ -600,7 +600,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         self.configure_kv_cache_dtype()
 
         # Init memory pool and attention backends
-        self.init_memory_pool(min_per_gpu_memory)
+        self.init_memory_pool(pre_model_load_memory)
 
         # Init max running requests
         self.max_running_requests = min(
@@ -826,7 +826,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             if is_npu():
                 register_sgl_tp_rank(self.gpu_id)
 
-        min_per_gpu_memory = get_available_gpu_memory(
+        pre_model_load_memory = get_available_gpu_memory(
             self.device,
             self.gpu_id,
             distributed=get_world_group().world_size > 1,
@@ -839,9 +839,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         # Check memory for tensor parallelism
         local_gpu_memory = get_available_gpu_memory(self.device, self.gpu_id)
         if self.tp_size > 1 and not self.is_draft_worker:
-            if min_per_gpu_memory < local_gpu_memory * 0.9:
+            if pre_model_load_memory < local_gpu_memory * 0.9:
                 msg = "The memory capacity is unbalanced. Some GPUs may be occupied by other processes. "
-                msg += f"{min_per_gpu_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
+                msg += f"{pre_model_load_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
                 if envs.SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK.get():
                     raise RuntimeError(msg)
                 else:
@@ -851,7 +851,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             f"Init torch distributed ends. elapsed={time.perf_counter() - tic:.2f} s, "
             f"mem usage={(before_avail_memory - local_gpu_memory):.2f} GB"
         )
-        return min_per_gpu_memory
+        return pre_model_load_memory
 
     def init_shared_mooncake_transfer_engine(self):
         """

python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py

@@ -113,8 +113,8 @@ class ModelRunnerKVCacheMixin:
                 )
         return cell_size
 
-    def profile_max_num_token(self: ModelRunner, total_gpu_memory: int):
-        available_gpu_memory = get_available_gpu_memory(
+    def profile_max_num_token(self: ModelRunner, pre_model_load_memory: int):
+        post_model_load_memory = get_available_gpu_memory(
             self.device,
             self.gpu_id,
             distributed=get_world_group().world_size > 1,
@@ -140,7 +140,7 @@ class ModelRunnerKVCacheMixin:
 
         cell_size = self.get_cell_size_per_token(num_layers)
 
-        rest_memory = available_gpu_memory - total_gpu_memory * (
+        rest_memory = post_model_load_memory - pre_model_load_memory * (
             1 - self.mem_fraction_static
         )
         if self.mambaish_config is not None:
@@ -675,10 +675,10 @@ class ModelRunnerKVCacheMixin:
                     self.token_to_kv_pool_allocator.full_to_swa_index_mapping
                 )
 
-    def init_memory_pool(self: ModelRunner, total_gpu_memory: int):
+    def init_memory_pool(self: ModelRunner, pre_model_load_memory: int):
         max_num_reqs = self.server_args.max_running_requests
         max_total_tokens = self.server_args.max_total_tokens
-        self.max_total_num_tokens = self.profile_max_num_token(total_gpu_memory)
+        self.max_total_num_tokens = self.profile_max_num_token(pre_model_load_memory)
 
         if max_num_reqs is None:
             max_num_reqs = min(