Skip to content
View puranikyashaswin's full-sized avatar
4 followers · 3 following

Achievements

Achievement: Pull Shark

Achievements

Achievement: Pull Shark

Block or report puranikyashaswin

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
puranikyashaswin/README.md

Puranik Yashaswin Sharma

I work on ML systems from two ends of the stack: Internals (Reverse-engineering representations, mechanistic interpretability) and Systems (Bare-metal inference, memory bandwidth optimization).

CS Undergrad @ Woxsen University | Core Contributor: TransformerLens & SAELens

The Stack

  • Internals: Activation patching, linear probing, causal circuit discovery, equation-to-code architecture replication.
  • Systems: C/C++, ARM NEON SIMD intrinsics, flat memory allocation, cache-line streaming.
  • Tools: PyTorch, wgpu/Metal, NetworkX.

Work

  • AutoCircuit Automated causal circuit discovery on GPT-2 Small. Scores all 156 components via activation patching. Found a minimal 3-component circuit recovering 93.4% of IOI performance. 90 pairwise passes vs. ~24,000 exhaustivea 99.6% reduction in search space.

  • bare-metal-transformer Transformer inference in pure C++. Zero ML dependencies (No PyTorch, no GGML). Flat KV cache via malloc, matmul with 16-wide ARM NEON FMA unrolling, and i-k-j loop ordering. 125 tok/s vs PyTorch's single-threaded 113 tok/s on TinyStories-33M.

  • gqa-inference-engine Equation-to-code replication of Grouped Query Attention (arXiv:2305.13245). Zero use of nn.MultiheadAttention. Handled causal mask shape mismatches during cached decoding natively. Verified 4x KV cache reduction (1048KB → 262KB) and 16x generation speedup.

  • gpt2-hallucination-probe Linear probing across all 12 residual stream layers of GPT-2 Small using a custom 46-prompt, 3-tier dataset. Peak AUROC 0.708 at Layer 0, dropping to chance on syntax-matched controls. Finding: The probe exploits surface lexical artifacts, not epistemic state. Documented as a legitimate negative result.

  • rusty Ongoing: GPU-accelerated ML training framework in pure Rust using wgpu/Metal.

Self-directed, shipping code, and always looking for the next hard bottleneck.

Pinned Loading

  1. AutoCircuit AutoCircuit Public

    Automated causal circuit discovery for GPT-2 Small. 99.6% more efficient than exhaustive ACDC, 3-component minimal circuit recovers 93.4% IOI behavior.

    Python

  2. bare-metal-transformer bare-metal-transformer Public

    Zero-dependency C++ transformer inference engine. ARM NEON SIMD, flat KV cache via malloc. 125 tok/s vs PyTorch 113 tok/s on TinyStories-33M.

    C++

  3. gpt2-hallucination-probe gpt2-hallucination-probe Public

    Linear probing suite on GPT-2 residual stream across all 12 layers. Documented negative result probe exploits lexical artifacts, not true hallucination mechanism.

    Python

  4. gqa-inference-engine gqa-inference-engine Public

    GQA paper replication from scratch (arXiv:2305.13245). 4x KV cache reduction, 16x generation speedup. Zero use of nn.MultiheadAttention.

    Python

  5. rusty rusty Public

    🦀 GPU-Accelerated ML Training Framework in Pure Rust - wgpu/Metal

    Rust 2