kmabeeTT

#dram = #ttnn.buffer_type<dram>
#l1 = #ttnn.buffer_type<l1>
#loc = loc(unknown)
#loc67 = loc("p0.1")
#loc68 = loc("p1.5")
#loc69 = loc("p2.9")
#loc70 = loc("p3.14")
#loc71 = loc("p4.21")
#loc72 = loc("p5.27")
#loc73 = loc("p6.71")

kmabeeTT

TT-XLA: Python-Level Host Memory Leak After TP LLM Tests

Date: 2026-03-01 Repo: tt-xla PR: #3521

---

Problem

kmabeeTT

Claude Code Sound Notification Hooks

Audio alerts for when Claude finishes a task or needs your permission.

Setup

Add this hooks block to ~/.claude/settings.json:

{

kmabeeTT

================================================================================
Test 1: Exact boundary (Wt in tiles)
================================================================================
  [PASS] Wt=  1: size=32, unique=32, dups=0, worst=0
  [PASS] Wt=  2: size=64, unique=64, dups=0, worst=0
  [PASS] Wt=  4: size=128, unique=128, dups=0, worst=0
  [PASS] Wt=  8: size=256, unique=256, dups=0, worst=0
  [FAIL] Wt=  9: size=288, unique=287, dups=1, worst=2
  [FAIL] Wt= 10: size=320, unique=319, dups=1, worst=2
  [FAIL] Wt= 11: size=352, unique=351, dups=1, worst=2

kmabeeTT

#loc1 = loc("p0.1")
#loc2 = loc("p1.37")
#loc3 = loc("p2.42")
#loc4 = loc("p3.82")
#loc5 = loc("p4.121")
#loc6 = loc("p5.150")
module @SyncTensorsGraph.296 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
  ttcore.device_module {
    builtin.module @SyncTensorsGraph.296 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
      func.func @main(%arg0: tensor<i64> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<i64>>} loc("p0.1"), %arg1: tensor<1xf32> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape =

kmabeeTT

#loc1 = loc("p0.1")
module @SyncTensorsGraph.25 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
  ttcore.device_module {
    builtin.module @SyncTensorsGraph.25 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
      func.func @main(%arg0: tensor<1x50000xf32> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<1x50000xf32>>, ttir.name = "args_0"} loc("p0.1")) -> (tensor<1x100xi64> {ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<1x100xi64>>}) {
        %0 = "ttir.reshape"(%arg0) <{shape = [1 : i32, 1 : i32, 50000 : i32]}> : (tensor<1x50000xf3

kmabeeTT

ttnn.sort UINT16/UINT32 Indices Type Mismatch

Summary

• Test: topk with input size (1, 50000)
• Graph: 1
• Operation: ttnn.sort at line 9
• Error: DataType mismatch, expected UINT16, got UINT32
• Scope: Debug builds only (release builds pass but have latent bug)

Root Cause

kmabeeTT

Simple Repro Commands

These commands reproduce the ttnn.sort f32 type mismatch failure.

Prerequisites

Ensure you have:

• opt125m_graph_17_ttir.mlir.txt (already in this directory)
• ttrt-artifacts/system_desc.ttsys (in parent directory)

kmabeeTT

#loc1 = loc("p0.1")
#loc2 = loc("p1.37")
#loc3 = loc("p2.42")
#loc4 = loc("p3.82")
#loc5 = loc("p4.121")
#loc6 = loc("p5.150")
module @SyncTensorsGraph.296 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
  ttcore.device_module {
    builtin.module @SyncTensorsGraph.296 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
      func.func @main(%arg0: tensor<i64> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<i64>>} loc("p0.1"), %arg1: tensor<1xf32> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape =

kmabeeTT

#loc1 = loc("p0.1")
module @SyncTensorsGraph.25 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
  ttcore.device_module {
    builtin.module @SyncTensorsGraph.25 attributes {mhlo.cross_program_prefetches = [], mhlo.input_output_alias = [], mhlo.is_dynamic = false, mhlo.use_auto_spmd_partitioning = false, ttcore.meshes = #ttcore.meshes<[<"mesh" = 1x1>]>} {
      func.func @main(%arg0: tensor<1x10xf32> {ttcore.argument_type = #ttcore.argument_type<input>, ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<1x10xf32>>, ttir.name = "args_0"} loc("p0.1")) -> (tensor<1x5xi64> {ttcore.runtime_tensor_sharding = #ttcore<runtime_tensor_sharding shard_status = <unsharded>, local_shape = tensor<1x5xi64>>}) {
        %0 = "ttir.reshape"(%arg0) <{shape = [1 : i32, 1 : i32, 10 : i32]}> : (tensor<1x10xf32>) -> tensor<1x