refactor: create functions for shard/tensor size calculations (#3257)

Summary:
Pull Request resolved: #3257

- refactor to create function for checking if table is cached
- refactor to create functions for tensor size calculations

Reviewed By: levythu, aporialiao

Differential Revision: D79007077

fbshipit-source-id: 9c514bd86cbef9f85cfb10238b2ba294d344d12b

Author: Ahmed Shuaibi

torchrec/distributed/planner/shard_estimators.py

@@ -1197,16 +1197,9 @@ def calculate_shard_storages(
     hbm_storage: int = tensor_storage.get("hbm", 0)
     ddr_storage: int = tensor_storage.get("ddr", 0)
 
-    table_cached: bool = False
-    if compute_kernel in {
-        EmbeddingComputeKernel.FUSED_UVM_CACHING.value,
-        EmbeddingComputeKernel.QUANT_UVM_CACHING.value,
-        EmbeddingComputeKernel.KEY_VALUE.value,
-        EmbeddingComputeKernel.SSD_VIRTUAL_TABLE.value,
-        EmbeddingComputeKernel.DRAM_VIRTUAL_TABLE.value,
-    }:
+    table_cached = _is_table_cached(compute_kernel)
+    if table_cached:
         hbm_storage = round(ddr_storage * caching_ratio)
-        table_cached = True
 
     optimizer_class = getattr(tensor, "_optimizer_classes", [None])[0]
 
@@ -1304,6 +1297,20 @@ def calculate_shard_storages(
     ]
 
 
+def _is_table_cached(
+    compute_kernel: str,
+) -> bool:
+    if compute_kernel in {
+        EmbeddingComputeKernel.FUSED_UVM_CACHING.value,
+        EmbeddingComputeKernel.QUANT_UVM_CACHING.value,
+        EmbeddingComputeKernel.KEY_VALUE.value,
+        EmbeddingComputeKernel.SSD_VIRTUAL_TABLE.value,
+        EmbeddingComputeKernel.DRAM_VIRTUAL_TABLE.value,
+    }:
+        return True
+    return False
+
+
 def _calculate_shard_io_sizes(
     sharding_type: str,
     batch_sizes: List[int],
@@ -1565,27 +1572,20 @@ def _calculate_storage_specific_sizes(
     is_inference: bool = False,
     clf: Optional[float] = None,
 ) -> List[int]:
-    tensor_sizes: List[int] = [
-        (
-            math.ceil(storage * prod(size) / prod(shape))
-            if sharding_type != ShardingType.DATA_PARALLEL.value
-            else storage
-        )
-        for size in shard_sizes
-    ]
-    optimizer_multipler: float = _get_optimizer_multipler(optimizer_class, shape)
-
-    optimizer_sizes: List[int] = [
-        math.ceil(tensor_size * optimizer_multipler) for tensor_size in tensor_sizes
-    ]
-
-    # If a table has turned on UVM caching (meaning clf is not None), there'll be
-    # 4x of table hash size and 16x of cache slot size HBM storage cost dedicated to
-    # cache aux state (note that this is not the cache content itself)
-    cache_aux_state_sizes: List[int] = (
-        [0] * len(shard_sizes)
-        if clf is None
-        else [math.ceil(size[0] * (4 + clf * 16)) for size in shard_sizes]
+    tensor_sizes: List[int] = _calculate_tensor_sizes(
+        storage,
+        shape,
+        shard_sizes,
+        sharding_type,
+    )
+    optimizer_sizes = _calculate_optimizer_sizes(
+        tensor_sizes,
+        optimizer_class,
+        shape,
+    )
+    cache_aux_state_sizes: List[int] = _calculate_cache_aux_state_sizes(
+        shard_sizes,
+        clf,
     )
 
     return [
@@ -1600,6 +1600,45 @@ def _calculate_storage_specific_sizes(
     ]
 
 
+def _calculate_tensor_sizes(
+    storage: int, shape: torch.Size, shard_sizes: List[List[int]], sharding_type: str
+) -> List[int]:
+    return [
+        (
+            math.ceil(storage * prod(size) / prod(shape))
+            if sharding_type != ShardingType.DATA_PARALLEL.value
+            else storage
+        )
+        for size in shard_sizes
+    ]
+
+
+# If a table has turned on UVM caching (meaning clf is not None), there'll be
+# 4x of table hash size and 16x of cache slot size HBM storage cost dedicated to
+# cache aux state (note that this is not the cache content itself)
+def _calculate_cache_aux_state_sizes(
+    shard_sizes: List[List[int]], clf: Optional[float]
+) -> List[int]:
+    if clf is None:
+        return [0] * len(shard_sizes)
+    return [math.ceil(size[0] * (4 + clf * 16)) for size in shard_sizes]
+
+
+def _calculate_optimizer_sizes(
+    tensor_sizes: List[int],
+    optimizer_class: Optional[Type[torch.optim.Optimizer]],
+    sharding_tensor_shape: torch.Size,
+) -> List[int]:
+    optimizer_multiplier: float = _get_optimizer_multipler(
+        optimizer_class,
+        sharding_tensor_shape,
+    )
+    optimizer_sizes: List[int] = [
+        math.ceil(tensor_size * optimizer_multiplier) for tensor_size in tensor_sizes
+    ]
+    return optimizer_sizes
+
+
 def _get_optimizer_multipler(
     optimizer_class: Optional[Type[torch.optim.Optimizer]],
     shape: torch.Size,