Per-weight constant cache for CUDA backend

Replace the old update_constants_from_blob + cross-method sharing with a
unified per-weight caching approach. The first method to initialize loads
its constants from the blob and caches them by FQN. Subsequent methods
with matching FQNs reuse cached GPU tensors via
update_user_managed_constant_buffer_pairs, skipping blob loading entirely.

This eliminates duplicate GPU weight allocations for multi-method models
(e.g., prefill/decode), reducing peak GPU memory from ~35 GB to ~17.6 GB
for Qwen 3.5 MoE.

Also adds GPU peak memory reporting to the Qwen3.5 MoE runner and a
CI check (< 20 GB) in test_model_e2e.sh.

Author: Gasoonjia

backends/aoti/aoti_delegate_handle.h

@@ -104,23 +104,6 @@ using AOTInductorModelContainerGetConstantOriginalFQNFunc =
         size_t idx,
         const char** original_fqn);
 
-// Retrieves a constant's data size in bytes by index.
-using AOTInductorModelContainerGetConstantDataSizeFunc = AOTIRuntimeError (*)(
-    AOTInductorModelContainerHandle container_handle,
-    size_t idx,
-    size_t* data_size);
-
-// Retrieves whether a constant was produced by constant folding.
-using AOTInductorModelContainerGetConstantFromFoldedFunc = AOTIRuntimeError (*)(
-    AOTInductorModelContainerHandle container_handle,
-    size_t idx,
-    bool* from_folded);
-
-// Retrieves the total size of the constants blob.
-using AOTInductorModelContainerGetConstantsBlobSizeFunc = AOTIRuntimeError (*)(
-    AOTInductorModelContainerHandle container_handle,
-    uint64_t* ret_size);
-
 // Extracts the constants map from the container (active or inactive buffer).
 // constant_map_handle should point to a
 // std::unordered_map<std::string, AtenTensorHandle>.
@@ -160,9 +143,6 @@ struct AOTIDelegateHandle {
   AOTInductorModelContainerGetNumConstantsFunc get_num_constants;
   AOTInductorModelContainerGetConstantNameFunc get_constant_name;
   AOTInductorModelContainerGetConstantOriginalFQNFunc get_constant_original_fqn;
-  AOTInductorModelContainerGetConstantDataSizeFunc get_constant_data_size;
-  AOTInductorModelContainerGetConstantFromFoldedFunc get_constant_from_folded;
-  AOTInductorModelContainerGetConstantsBlobSizeFunc get_constants_blob_size;
   AOTInductorModelContainerExtractConstantsMapFunc extract_constants_map;
   AOTInductorModelContainerUpdateUserManagedConstantBufferPairsFunc
       update_user_managed_constant_buffer_pairs;

backends/arm/_passes/__init__.py

@@ -53,6 +53,7 @@
 from .decompose_glu_pass import DecomposeGluPass  # noqa
 from .decompose_grouped_conv_pass import DecomposeGroupedConvPass  # noqa
 from .decompose_groupnorm_pass import DecomposeGroupNormPass  # noqa
+from .decompose_gru_pass import DecomposeGruPass  # noqa
 from .decompose_index_copy_pass import DecomposeIndexCopyPass  # noqa
 from .decompose_index_select_to_gather_pass import (  # noqa
     DecomposeIndexSelectToGatherPass,
@@ -70,13 +71,15 @@
 from .decompose_linear_pass import DecomposeLinearPass  # noqa
 from .decompose_log1p_pass import DecomposeLog1pPass  # noqa
 from .decompose_logit_pass import DecomposeLogitPass  # noqa
+from .decompose_lstm_pass import DecomposeLstmPass  # noqa
 from .decompose_masked_fill_pass import DecomposeMaskedFillPass  # noqa
 from .decompose_matmul import DecomposeMatmulPass  # noqa
 from .decompose_maxpool2d_with_dilation_pass import DecomposeMaxPool2dPass  # noqa
 from .decompose_meandim_pass import DecomposeMeanDimPass  # noqa
 from .decompose_ne_pass import DecomposeNotEqualPass  # noqa
 from .decompose_quant_nodes import DecomposeQuantNodesPass  # noqa
 from .decompose_remainder_pass import DecomposeRemainderPass  # noqa
+from .decompose_rnn_pass import DecomposeRnnPass  # noqa
 from .decompose_round_pass import DecomposeRoundPass  # noqa
 from .decompose_sdpa_pass import DecomposeScaledDotProductAttentionPass  # noqa
 from .decompose_select import DecomposeSelectPass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -61,6 +61,7 @@
     DecomposeGluPass,
     DecomposeGroupedConvPass,
     DecomposeGroupNormPass,
+    DecomposeGruPass,
     DecomposeIndexCopyPass,
     DecomposeIndexSelectToGatherPass,
     DecomposeIndexTensorToGatherPass,
@@ -71,13 +72,15 @@
     DecomposeLinearPass,
     DecomposeLog1pPass,
     DecomposeLogitPass,
+    DecomposeLstmPass,
     DecomposeMaskedFillPass,
     DecomposeMatmulPass,
     DecomposeMaxPool2dPass,
     DecomposeMeanDimPass,
     DecomposeNotEqualPass,
     DecomposeQuantNodesPass,
     DecomposeRemainderPass,
+    DecomposeRnnPass,
     DecomposeRoundPass,
     DecomposeScaledDotProductAttentionPass,
     DecomposeSelectPass,
@@ -360,6 +363,9 @@ def _tosa_pipeline(
                 ConvertToClampPass(),
                 DecomposeTOSAUnsupportedClampPass(),
                 DecomposeGroupNormPass(),
+                DecomposeGruPass(),
+                DecomposeLstmPass(),
+                DecomposeRnnPass(),
                 DecomposeLayerNormPass(),
                 DecomposeVarPass(),
                 DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec),
@@ -578,6 +584,9 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_passes(
             [
                 NormalizeWhileInitialArgsPass(use_exir_clone=False, tfa_pass=True),
+                DecomposeGruPass(tfa_pass=True),
+                DecomposeLstmPass(tfa_pass=True),
+                DecomposeRnnPass(tfa_pass=True),
                 DecomposeNotEqualPass(tfa_pass=True),
                 DecomposeCosineSimilarityPass(tfa_pass=True),
                 DecomposeGluPass(tfa_pass=True),