tenstorrent/tt-mlir
Jul 2025 – Mar 2026
9 Months active
Languages Used
C++MLIRPythonShell
Technical Skills
Compiler DevelopmentLinear AlgebraMLIRTensor OperationsCompiler designGPU programming
Djordje Stefanovic
PROFILE
Djordje Stefanovic
During his nine-month tenure, Dragan Stefanovic engineered core compiler and backend infrastructure for the tenstorrent/tt-mlir repository, focusing on MLIR-based lowering, operator fusion, and backend portability. He developed and optimized conversion pipelines for TTIR and TTNN dialects, enabling efficient execution on both CPU and GPU targets. Using C++, MLIR, and Python, Dragan implemented advanced features such as Conv2D/Conv3D lowering, operator fusion for GELU and SILU, and memory optimization passes to reduce DRAM pressure. His work addressed correctness, performance, and test coverage, resulting in robust support for deep learning workloads and improved reliability for production machine learning deployments.
Overall Statistics
Feature vs Bugs
75%Features
Repository Contributions
53Total
Bugs
9
Commits
53
Features
27
Lines of code
21,781
Activity Months9
Your Network
438 peopleWork History
March 2026
8 Commits • 3 Features
Mar 1, 2026Month: 2026-03. Focused on stability, correctness, and efficiency for the tenstorrent/tt-mlir project. Key features delivered include Conv3D reliability cleanup with test coverage and simplification of rewrite patterns; group normalization support across TTIR and TTNN; and a memory optimization pass to reduce DRAM pressure. Major bugs fixed include negative bound indexing in slice operations, non tile-aligned inner dimension padding for ttnn.prod, a UINT8 comparison workaround, and a safety check for zero-volume dimensions in affine maps. Builder tests and conversion patterns increased coverage, reducing regression risk. Overall impact: improved ML model deployment reliability, lower memory footprint, and broader op coverage, enabling more robust TTIR/TTNN workflows. Technologies/skills demonstrated: MLIR dialect work (TTIR/TTNN), rewrite pattern refinement, pattern-driven optimization, builder/test-driven development, and memory-optimization techniques to boost efficiency.
8 Commits • 3 Features
Mar 1, 2026Month: 2026-03. Focused on stability, correctness, and efficiency for the tenstorrent/tt-mlir project. Key features delivered include Conv3D reliability cleanup with test coverage and simplification of rewrite patterns; group normalization support across TTIR and TTNN; and a memory optimization pass to reduce DRAM pressure. Major bugs fixed include negative bound indexing in slice operations, non tile-aligned inner dimension padding for ttnn.prod, a UINT8 comparison workaround, and a safety check for zero-volume dimensions in affine maps. Builder tests and conversion patterns increased coverage, reducing regression risk. Overall impact: improved ML model deployment reliability, lower memory footprint, and broader op coverage, enabling more robust TTIR/TTNN workflows. Technologies/skills demonstrated: MLIR dialect work (TTIR/TTNN), rewrite pattern refinement, pattern-driven optimization, builder/test-driven development, and memory-optimization techniques to boost efficiency.
March 2026
February 2026
7 Commits • 3 Features
Feb 1, 2026February 2026 monthly summary for tenstorrent/tt-mlir. Delivered key MLIR-based enhancements and reliability improvements across Conv3D, padding, and memory configuration, with a focus on business value and robust cross-framework support.
February 2026
7 Commits • 3 Features
Feb 1, 2026February 2026 monthly summary for tenstorrent/tt-mlir. Delivered key MLIR-based enhancements and reliability improvements across Conv3D, padding, and memory configuration, with a focus on business value and robust cross-framework support.
January 2026
10 Commits • 4 Features
Jan 1, 2026January 2026 monthly summary for tenstorrent/tt-mlir: Delivered substantial TTIR/TTNN enhancements, expanded operator fusion and normalization coverage, and EmitC generation improvements. The work improved performance, memory efficiency, and correctness while broadening operator coverage and configurability, enabling faster compilation and more scalable deployments.
10 Commits • 4 Features
Jan 1, 2026January 2026 monthly summary for tenstorrent/tt-mlir: Delivered substantial TTIR/TTNN enhancements, expanded operator fusion and normalization coverage, and EmitC generation improvements. The work improved performance, memory efficiency, and correctness while broadening operator coverage and configurability, enabling faster compilation and more scalable deployments.
January 2026
December 2025
11 Commits • 9 Features
Dec 1, 2025December 2025 (tenstorrent/tt-mlir) delivered a set of business-critical features that improve LLM readiness, runtime performance, and stability, while strengthening test coverage and API ergonomics. Notable outcomes include LLM backprop support via GELU gradient integration, performance gains through SILU matmul fusion, and StableHLO compatibility improvements through pooling reshaping; plus memory and configurability enhancements for deployment reliability.
December 2025
11 Commits • 9 Features
Dec 1, 2025December 2025 (tenstorrent/tt-mlir) delivered a set of business-critical features that improve LLM readiness, runtime performance, and stability, while strengthening test coverage and API ergonomics. Notable outcomes include LLM backprop support via GELU gradient integration, performance gains through SILU matmul fusion, and StableHLO compatibility improvements through pooling reshaping; plus memory and configurability enhancements for deployment reliability.
November 2025
7 Commits • 2 Features
Nov 1, 2025November 2025 monthly summary for tenstorrent/tt-mlir focusing on business value and technical achievements. Key features delivered center on stabilizing and unifying the convolution path in TTIR with StableHLO integration, while improvements to code generation and test coverage reinforce reliability for production workloads.
7 Commits • 2 Features
Nov 1, 2025November 2025 monthly summary for tenstorrent/tt-mlir focusing on business value and technical achievements. Key features delivered center on stabilizing and unifying the convolution path in TTIR with StableHLO integration, while improvements to code generation and test coverage reinforce reliability for production workloads.
November 2025
October 2025
5 Commits • 3 Features
Oct 1, 2025In October 2025, delivered key TTIR/TTNN pipeline enhancements and feature implementations across Conv2d lowering, GPU pipeline optimizations, and StableHLO compatibility, with comprehensive tests and clear business value for production readiness on GPU accelerators.
October 2025
5 Commits • 3 Features
Oct 1, 2025In October 2025, delivered key TTIR/TTNN pipeline enhancements and feature implementations across Conv2d lowering, GPU pipeline optimizations, and StableHLO compatibility, with comprehensive tests and clear business value for production readiness on GPU accelerators.
September 2025
2 Commits • 1 Features
Sep 1, 2025September 2025 monthly summary for tenstorrent/tt-mlir focused on expanding TTIR operator support, strengthening correctness, and enabling backend mappings to TOSA/Linalg with robust test coverage. Highlights: - Delivered TTIR to Linalg conversion improvements for squeeze and mean, enabling broader operator support. Implemented squeeze via tosa.reshape and mean via tosa.reduce_sum followed by linalg.div; includes test coverage. - Implemented MaxPool2d lowering from TTIR to TOSA with strict dimension checks, padding handling, and final slicing, backed by tests. - Fixed critical bugs: ttir.reluOp now supports bf16 element types, and createReductionOpChain correctly handles keepDim = false. - Comprehensive test coverage accompanies all changes to ensure regression safety and future refactorability. Impact: - Business value: expands model compatibility and reduces manual lowering work, accelerating feature delivery and reliability for downstream ML workloads. - Technical achievements: end-to-end TTIR->Linalg/TOSA conversion improvements, stronger correctness guarantees, and improved backend interoperability. Technologies/skills demonstrated: - MLIR TTIR, Linalg and TOSA dialect interactions - Lowering/pattern matching, dimension/padding logistics, and slicing logic - bf16 support and reduction ops, test-driven development
2 Commits • 1 Features
Sep 1, 2025September 2025 monthly summary for tenstorrent/tt-mlir focused on expanding TTIR operator support, strengthening correctness, and enabling backend mappings to TOSA/Linalg with robust test coverage. Highlights: - Delivered TTIR to Linalg conversion improvements for squeeze and mean, enabling broader operator support. Implemented squeeze via tosa.reshape and mean via tosa.reduce_sum followed by linalg.div; includes test coverage. - Implemented MaxPool2d lowering from TTIR to TOSA with strict dimension checks, padding handling, and final slicing, backed by tests. - Fixed critical bugs: ttir.reluOp now supports bf16 element types, and createReductionOpChain correctly handles keepDim = false. - Comprehensive test coverage accompanies all changes to ensure regression safety and future refactorability. Impact: - Business value: expands model compatibility and reduces manual lowering work, accelerating feature delivery and reliability for downstream ML workloads. - Technical achievements: end-to-end TTIR->Linalg/TOSA conversion improvements, stronger correctness guarantees, and improved backend interoperability. Technologies/skills demonstrated: - MLIR TTIR, Linalg and TOSA dialect interactions - Lowering/pattern matching, dimension/padding logistics, and slicing logic - bf16 support and reduction ops, test-driven development
September 2025
August 2025
1 Commits • 1 Features
Aug 1, 2025Summary for August 2025: Delivered a GPU execution path for TTIR by implementing a TTIR-to-NVVM/LLVM pipeline in tenstorrent/tt-mlir, enabling TTIR workloads to run on NVIDIA GPUs. The work includes converting TTIR to NVVM and lowering to PTX, with logic to wrap single AffineFor loops to generate correct GPU kernels. Enhanced the MLIR framework to support GPU optimizations and conversions, strengthening cross-backend portability and performance. This milestone expands execution targets, enabling GPU-accelerated workloads and laying the foundation for future performance gains and broader hardware support.
August 2025
1 Commits • 1 Features
Aug 1, 2025Summary for August 2025: Delivered a GPU execution path for TTIR by implementing a TTIR-to-NVVM/LLVM pipeline in tenstorrent/tt-mlir, enabling TTIR workloads to run on NVIDIA GPUs. The work includes converting TTIR to NVVM and lowering to PTX, with logic to wrap single AffineFor loops to generate correct GPU kernels. Enhanced the MLIR framework to support GPU optimizations and conversions, strengthening cross-backend portability and performance. This milestone expands execution targets, enabling GPU-accelerated workloads and laying the foundation for future performance gains and broader hardware support.
July 2025
2 Commits • 1 Features
Jul 1, 2025July 2025 focused on strengthening TTIR lowering paths in tenstorrent/tt-mlir and improving correctness of the LINALG/TOSA backends. Key feature delivered: TTIR.relu Support in LINALG lowering, implemented via a zero-tensor broadcast with linalg.max to reproduce ReLU behavior and ensure correct lowering of TTIR.relu to the LINALG dialect (commit 0c40d8385abd614adcf0c0f7523deb7074aa14a5). Major bug fixes: TTIR lowering fixes to LINALG and TOSA, including corrected matmul reshaping and added broadcasting for ttir.add with differing input ranks; tests added to cover these fixes (commit c502a663baf51f911dad6995d55d39a5cef5e7d8). Overall impact: increases reliability and readiness of the lowering path for real-world models, reduces risk of regressions, and expands test coverage to guard critical linear algebra ops. Technologies demonstrated: MLIR TTIR dialect lowering, LINALG and TOSA backends, tensor broadcasting, zero-tensor based ReLU, and test-driven development.
2 Commits • 1 Features
Jul 1, 2025July 2025 focused on strengthening TTIR lowering paths in tenstorrent/tt-mlir and improving correctness of the LINALG/TOSA backends. Key feature delivered: TTIR.relu Support in LINALG lowering, implemented via a zero-tensor broadcast with linalg.max to reproduce ReLU behavior and ensure correct lowering of TTIR.relu to the LINALG dialect (commit 0c40d8385abd614adcf0c0f7523deb7074aa14a5). Major bug fixes: TTIR lowering fixes to LINALG and TOSA, including corrected matmul reshaping and added broadcasting for ttir.add with differing input ranks; tests added to cover these fixes (commit c502a663baf51f911dad6995d55d39a5cef5e7d8). Overall impact: increases reliability and readiness of the lowering path for real-world models, reduces risk of regressions, and expands test coverage to guard critical linear algebra ops. Technologies demonstrated: MLIR TTIR dialect lowering, LINALG and TOSA backends, tensor broadcasting, zero-tensor based ReLU, and test-driven development.
July 2025
Activity
Loading activity data...
Quality Metrics
Correctness95.2%
Maintainability82.6%
Architecture87.8%
Performance82.6%
AI Usage28.0%
Skills & Technologies
Programming Languages
C++MLIRPythonShell
Technical Skills
Algorithm OptimizationBackend DevelopmentBug FixingC++C++ DevelopmentC++ ProgrammingC++ developmentCode ConversionCode GenerationCode OptimizationCompiler DesignCompiler DevelopmentCompiler designConvolutional Neural NetworksDeep Learning Operations
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline