Per weight constant cache (#18901)

**Problem**: Multi-method AOTI models (e.g., Qwen3.5 MoE with separate
prefill/decode methods) load the full weight blob independently for each
method, even when they share identical weights. This causes duplicate
GPU allocations -- Qwen3.5 MoE peaked at ~35 GB, making it impossible to
run on a single 24 GB GPU (e.g., 4090).

**Solution**: Introduce a per-weight FQN-keyed constant cache in
`CudaBackend`. The first method loads its constants from the blob and
caches them. Subsequent methods with matching FQNs skip blob loading
entirely and reuse cached GPU tensors via
`update_user_managed_constant_buffer_pairs`. A legacy fallback path is
preserved for older AOTI models without constant management APIs.

**Results**
Peak GPU memory: 35.4 GB → 17.6 GB (-50%)

---------

Co-authored-by: gasoonjia <gasoonjia@fb.com>

Author: Gasoonjia

.ci/scripts/test_model_e2e.sh

@@ -397,6 +397,27 @@ if [ -n "$EXPECTED_OUTPUT" ]; then
 else
   echo "SUCCESS: Runner completed successfully"
 fi
+
+# Validate GPU peak memory usage for models with known memory budgets.
+# The runner prints "GPU peak memory usage: XXXX.X MiB" at the end.
+case "$MODEL_NAME" in
+  qwen3_5_moe)
+    MAX_MEMORY_MIB=20480  # 20 GB — must fit on a single GPU (e.g. 4090)
+    PEAK_MEM=$(echo "$OUTPUT" | grep -oP 'GPU peak memory usage: \K[0-9.]+' || true)
+    if [ -n "$PEAK_MEM" ]; then
+      # Compare as integers (truncate decimals)
+      PEAK_MEM_INT=${PEAK_MEM%%.*}
+      if [ "$PEAK_MEM_INT" -gt "$MAX_MEMORY_MIB" ]; then
+        echo "FAIL: GPU peak memory ${PEAK_MEM} MiB exceeds budget ${MAX_MEMORY_MIB} MiB"
+        exit 1
+      else
+        echo "Success: GPU peak memory ${PEAK_MEM} MiB within budget (max ${MAX_MEMORY_MIB} MiB)"
+      fi
+    else
+      echo "WARNING: GPU peak memory usage not found in output"
+    fi
+    ;;
+esac
 echo "::endgroup::"
 
 popd

backends/aoti/aoti_backend.py

@@ -25,7 +25,6 @@
 
 class COMPILE_SPEC_KEYS(Enum):
     METHOD_NAME = "method_name"
-    SHARE_KV_CACHE_ACROSS_METHODS = "share_kv_cache_across_methods"
 
 
 @experimental(
@@ -287,13 +286,3 @@ def method_name_from_compile_specs(
         raise RuntimeError(
             f"Could not find method name in compile specs: {compile_specs}"
         )
-
-    @classmethod
-    def generate_share_kv_cache_compile_spec(cls) -> CompileSpec:
-        """
-        Generate a CompileSpec to enable cross-method KV cache sharing.
-        """
-        return CompileSpec(
-            COMPILE_SPEC_KEYS.SHARE_KV_CACHE_ACROSS_METHODS.value,
-            bytes([1]),
-        )