Activity

1. github-actions

   added

   Office HoursIssue announcing office hours and soliciting associated agenda items.Issue announcing office hours and soliciting associated agenda items.

   on Oct 6, 2026
2. kgryte commented on Oct 12, 2026

   

   kgryte

   on Oct 12, 2026

   Member

   More actions

   This week's office hours will be held two hours earlier than usual, and I have updated the OP above accordingly.

   👍React with 👍5Reacted by Ujjwal Verma, Samarth Kolarkar, Aryan Sharma, Gunj Joshi and Chanho Lee
3. AryanSharma48 commented on Oct 13, 2026

   

   AryanSharma48

   on Oct 13, 2026

   More actions

   A few discussion points from my side:

   • Loss functions update: Synced with nakul and opened initial forward-pass PRs (e.g. squared-error) matching the existing gradient architecture.
   • Float32 ML pipeline scoping: Audited missing float32 packages across the monorepo; proposing a scoped tracking issue for ML loss & activation primitives while scoping out the Remez/minimax math backlog.
   • GELU & SiLU proposals: Finalizing namespace and numerical stability designs (erfc formulation, gelu.tanh, gelu.quick, and math/base/special/silu).
4. Erennn7 commented on Oct 13, 2026

   

   Erennn7

   on Oct 13, 2026

   More actions

   PR review on the C implementation for #16105
5. batpigandme commented on Oct 13, 2026

   

   batpigandme

   on Oct 13, 2026

   Last edited by batpigandme

   Member

   More actions

   AI-Generated Summary

   Caution

   These notes are likely imperfect, especially when it comes to spelling/function/package names

   Meeting Purpose

   Review open PRs and align on API design and implementation strategy.

   Key Takeaways

   • FFT PR #16089 requires a full refactor. The current implementation is 1D-only and inefficient, using manual buffer manipulation instead of the ndarray API's batching and optimized kernels.
   • Scalar parameters for native C implementations must be typed. To ensure safe, direct calls to native code, scalar parameters like k in ndarray-ctril must be explicitly typed as Int32Array, not ndarray-generic.
   • GELU approximations will be split into separate packages. To maintain a clean API, specialized approximations (tanh, quik) will live in their own packages (e.g., gelu-tanh) within the mathbase-special/fast namespace.
   • FFT API naming will be standardized. The API will use FFT for complex and RFFT for real transforms, with IFFT and IRFFT for inverses, replacing the ambiguous F (forward) and B (backward) suffixes.

   Topics

   FFT Implementation Refactor (PR #16089)

   • Problem: The current implementation is 1D-only and uses manual buffer manipulation, which is inefficient and prevents batch processing over multi-dimensional arrays.
   • Solution: Refactor the PR to use the ndarray API as a "recipe" for stitching together operations. This leverages optimized kernels for batching and avoids manual loops.
   • Refactor Steps:
     1. Workspace: Allocate a 1D workspace array, then broadcast its batch dimensions to match the input array.
     2. Input Copy: Use ndarray-base/assign to copy the input data into a slice of the output array.
     3. Computation: Pass the reinterpreted output array and the broadcasted workspace array to the base.js kernel.
     4. Post-processing: Use ndarray operations (assign, fill, scale) for cleanup (e.g., zeroing imaginary components).
   • Blocker: The ndarray-base/scale package is needed for post-processing. Athan will review the existing PR for it to unblock this work.

   FFT API Naming Standardization

   • Problem: The current F (forward) and B (backward) suffixes are ambiguous and inconsistent with common conventions.
   • Decision: Standardize on explicit names for clarity.
     • Complex: FFT (forward), IFFT (inverse)
     • Real: RFFT (forward), IRFFT (inverse)
   • Rationale: This aligns with conventions in libraries like MATLAB and Julia.
   • Documentation: The PR description will summarize this decision and its rationale. A formal mapping will be added to the NumPy migration guide later.

   Handling Scalar Parameters in Native C Implementations

   • Problem: For native C implementations, passing a generic ndarray for a scalar parameter (e.g., k in ndarray-ctril) is unsafe and requires inefficient casting.
   • Solution: Explicitly type scalar parameter arrays as Int32Array (or another appropriate integer type).
   • Rationale: This allows for direct, safe calls to native C functions without casting, improving performance and reliability.
   • Action: Yug will implement this fix for ndarray-ctril. Ujjwal will then help propagate the pattern to other affected PRs.

   GELU Package Structure

   • Problem: Multiple GELU approximations were planned for a single package, which creates an unclear API.
   • Decision: Split approximations into separate, named packages.
     • Base: gelu (textbook formula)
     • Approximations: gelu-tanh, gelu-quik (using the canonical quik name from VLLM/Mojo)
   • Location: These approximation packages will reside in the mathbase-special/fast sub-namespace to signal their performance-optimized nature, which may involve reduced accuracy.

   Full transcript

   Full office hours transcript

   ❤️React with ❤️1Reacted by Abdul Kaium