shunting314

#blocked = #ttg.blocked<{sizePerThread = [1, 8], threadsPerWarp = [4, 8], warpsPerCTA = [8, 1], order = [1, 0]}>
#blocked1 = #ttg.blocked<{sizePerThread = [1, 8], threadsPerWarp = [1, 32], warpsPerCTA = [8, 1], order = [1, 0]}>
#blocked2 = #ttg.blocked<{sizePerThread = [2, 1], threadsPerWarp = [32, 1], warpsPerCTA = [2, 4], order = [0, 1]}>
#linear = #ttg.linear<{register = [[0, 1], [0, 2], [0, 4], [0, 8], [0, 16], [0, 32], [0, 64]], lane = [[1, 0], [2, 0], [4, 0], [8, 0], [16, 0]], warp = [[32, 0], [64, 0], [0, 128]], block = []}>
#loc = loc("/home/shunting/ws/openai_triton/licm-for-inductor/standalone_bench.py":28:1)
#shared = #ttg.swizzled_shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [0]}>
#shared1 = #ttg.swizzled_shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [0, 1]}>
#shared2 = #ttg.nvmma_shared<{swizzlingByteWidth = 128, transposed = false, elementBitWidth = 16}>
#smem = #ttg.shared_memory
#tmem = #ttng.tensor_memory_encoding<blockM = 128, blockN = 256, colStride = 1>

shunting314

diff --git a/Makefile b/Makefile
index aff6ccd92c..0482f2aa9f 100644
--- a/Makefile
+++ b/Makefile
@@ -118,7 +118,7 @@ dev-install: dev-install-requires dev-install-triton
 .NOPARALLEL: dev-install-llvm
 dev-install-llvm:
 	LLVM_BUILD_PATH=$(LLVM_BUILD_PATH) scripts/build-llvm-project.sh
-	TRITON_BUILD_WITH_CLANG_LLD=1 TRITON_BUILD_WITH_CCACHE=0 \
+	TRITON_BUILD_WITH_CLANG_LLD=0 TRITON_BUILD_WITH_CCACHE=0 \

shunting314

diff --git a/scripts/build-llvm-project.sh b/scripts/build-llvm-project.sh
index e45356d5c3..795fed8149 100755
--- a/scripts/build-llvm-project.sh
+++ b/scripts/build-llvm-project.sh
@@ -18,9 +18,9 @@ if [ -z "$CMAKE_ARGS" ]; then
               -DCMAKE_BUILD_TYPE="$LLVM_BUILD_TYPE"
               -DLLVM_CCACHE_BUILD=OFF
               -DLLVM_ENABLE_ASSERTIONS=ON
-              -DCMAKE_C_COMPILER=clang
-              -DCMAKE_CXX_COMPILER=clang++

shunting314

diff --git a/scripts/build-llvm-project.sh b/scripts/build-llvm-project.sh
index e45356d5c3..dc65139895 100755
--- a/scripts/build-llvm-project.sh
+++ b/scripts/build-llvm-project.sh
@@ -18,8 +18,8 @@ if [ -z "$CMAKE_ARGS" ]; then
               -DCMAKE_BUILD_TYPE="$LLVM_BUILD_TYPE"
               -DLLVM_CCACHE_BUILD=OFF
               -DLLVM_ENABLE_ASSERTIONS=ON
-              -DCMAKE_C_COMPILER=clang
-              -DCMAKE_CXX_COMPILER=clang++

shunting314

 {
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA B200", "totalGlobalMem": 191503138816,
      "computeMajor": 10, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 148

shunting314

 {
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA B200", "totalGlobalMem": 191503138816,
      "computeMajor": 10, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 148

shunting314

[
  {
    "models": [
      "BAAI/bge-base-en-v1.5",
      "BAAI/bge-large-en-v1.5",
      "BAAI/bge-small-en-v1.5",
      "FinLang/finance-embeddings-investopedia",
      "ProsusAI/finbert",
      "Xenova/all-MiniLM-L6-v2",
      "colbert-ir/colbertv2.0",

shunting314

diff --git a/fbcode/helion/helion/__init__.py b/fbcode/helion/helion/__init__.py
--- a/fbcode/helion/helion/__init__.py
+++ b/fbcode/helion/helion/__init__.py
@@ -34,3 +34,6 @@
 from ._compiler._dynamo.variables import register_dynamo_variable  # noqa: E402
 
 register_dynamo_variable()
+
+import torch
+torch.cuda.memory._record_memory_history(max_entries=100000)

shunting314

Inductor:
        L_x_user: "f32[4096, 759, 128][97152, 128, 1]cuda:0",
        L_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_: "f32[192, 759][759, 1]cuda:0"
        l_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_ = L_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_
        l_x_user = L_x_user
        matmul: "bf16[4096, 192, 128][24576, 128, 1]cuda:0" = l_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_ @ l_x_user;  l_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_ = None

Helion
        l_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_ = L_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_
        L_self_modules_maybe_fuse

shunting314

--- /tmp/inductor.py	2026-04-30 10:22:24.247513427 -0700
+++ /tmp/helion.py	2026-04-30 10:22:40.475653971 -0700
@@ -1,8 +1,8 @@
 class GraphModule(torch.nn.Module):
-    def forward(self, L_x_ads: "f32[4096, 523, 128][66944, 128, 1]cuda:0", L_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_shared_parameters_weight_: "f32[192, 759][759, 1]cuda:0", L_x_user: "f32[4096, 759, 128][97152, 128, 1]cuda:0", L_x_ads_to_user_map: "i64[4096][1]cuda:0", L_self_modules_maybe_fused_lce_modules_lces_modules_0_modules_fc_modules_linear_ads_parameters_weight_: "f32[96, 523][523, 1]cuda:0", L_self_modules_maybe_fused_lce_modules_lces_modules_1_modules_fc_modules_linear_shared_parameters_weight_: "f32[32, 759][759, 1]cuda:0", L_self_modules_maybe_fused_lce_modules_lces_modules_1_modules_fc_modules_linear_ads_parameters_weight_: "f32[16, 523][523, 1]cuda:0", L_self_modules_maybe_fused_lce_modules_lces_modules_2_modules_fc_modules_linear_shared_parameters_weight_: "f32[192, 759][759, 1]cuda:0", L_sel