trnblas 0.2.0__tar.gz

trnblas-0.2.0/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,31 @@
+---
+name: Bug report
+about: Report a bug in trnblas
+labels: bug
+---
+
+## Environment
+
+- trnblas version:
+- Python version:
+- PyTorch version:
+- Hardware: CPU / trn1 / trn2 / inf2
+- OS:
+
+## Description
+
+What happened?
+
+## Steps to reproduce
+
+```python
+# Minimal code to reproduce
+```
+
+## Expected behavior
+
+What should have happened?
+
+## Actual behavior
+
+What happened instead? Include error messages/tracebacks if applicable.

trnblas-0.2.0/.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,17 @@
+---
+name: Feature request
+about: Suggest a new feature or enhancement
+labels: enhancement
+---
+
+## Description
+
+What feature would you like?
+
+## Motivation
+
+Why is this needed? What problem does it solve?
+
+## Alternatives considered
+
+Any other approaches you've thought about?

trnblas-0.2.0/.github/pull_request_template.md

@@ -0,0 +1,9 @@
+## Summary
+
+What does this PR do?
+
+## Test plan
+
+- [ ] Tests pass (`pytest tests/ -v`)
+- [ ] New tests added for new functionality
+- [ ] CHANGELOG.md updated (if user-facing change)

trnblas-0.2.0/.github/workflows/ci.yml

@@ -0,0 +1,20 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.10", "3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - run: pip install -e ".[dev]"
+      - run: pytest tests/ -v -x --tb=short -m "not neuron"

trnblas-0.2.0/.github/workflows/docs.yml

@@ -0,0 +1,38 @@
+name: Deploy Docs
+
+on:
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+  pages: write
+  id-token: write
+
+concurrency:
+  group: pages
+  cancel-in-progress: true
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install mkdocs mkdocs-material
+      - run: mkdocs build --strict
+      - uses: actions/upload-pages-artifact@v3
+        with:
+          path: site/
+
+  deploy:
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - id: deployment
+        uses: actions/deploy-pages@v4

trnblas-0.2.0/.github/workflows/neuron.yml

@@ -0,0 +1,50 @@
+name: Neuron Hardware Tests
+
+on:
+  workflow_dispatch:
+    inputs:
+      instance_type:
+        description: "Instance type"
+        required: true
+        default: "trn1"
+        type: choice
+        options:
+          - trn1
+          - trn2
+          - inf2
+
+env:
+  AWS_PROFILE: aws
+
+jobs:
+  neuron-test:
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: ${{ secrets.AWS_ROLE_ARN }}
+          aws-region: us-east-1
+
+      # TODO: Wire to actual Trainium instance via SSM or SSH.
+      # The steps below are a scaffold — replace with your instance access method.
+      #
+      # Example with SSM:
+      #   - name: Run tests on ${{ inputs.instance_type }}
+      #     run: |
+      #       INSTANCE_ID=$(aws ec2 describe-instances \
+      #         --filters "Name=tag:Name,Values=trnblas-ci-${{ inputs.instance_type }}" \
+      #         --query 'Reservations[0].Instances[0].InstanceId' --output text)
+      #       aws ssm send-command --instance-ids $INSTANCE_ID \
+      #         --document-name "AWS-RunShellScript" \
+      #         --parameters 'commands=["cd /home/ubuntu/trnblas && git pull && pip install -e \".[neuron,dev]\" && pytest tests/ -v -m neuron --tb=short"]'
+      - name: Run neuron tests on ${{ inputs.instance_type }}
+        run: |
+          echo "Neuron hardware CI scaffold"
+          echo "Instance type: ${{ inputs.instance_type }}"
+          echo "TODO: Wire to actual ${{ inputs.instance_type }} instance"
+          echo "Expected command: pytest tests/ -v -m neuron --tb=short"

trnblas-0.2.0/.github/workflows/publish.yml

@@ -0,0 +1,56 @@
+name: Publish to PyPI
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: pip install build
+      - run: python -m build
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish-pypi:
+    needs: build
+    runs-on: ubuntu-latest
+    # Only publish stable releases (v1.2.3), not pre-releases (v1.2.3-rc1)
+    if: ${{ !contains(github.event.release.tag_name, '-') }}
+    environment:
+      name: pypi
+      url: https://pypi.org/project/trnblas/
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1
+
+  publish-testpypi:
+    needs: build
+    runs-on: ubuntu-latest
+    # Only publish pre-releases to TestPyPI
+    if: ${{ contains(github.event.release.tag_name, '-') }}
+    environment:
+      name: testpypi
+      url: https://test.pypi.org/project/trnblas/
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://test.pypi.org/legacy/

trnblas-0.2.0/.gitignore

@@ -0,0 +1,18 @@
+__pycache__/
+*.py[cod]
+*$py.class
+*.egg-info/
+dist/
+build/
+.eggs/
+*.egg
+.pytest_cache/
+.tox/
+.venv/
+venv/
+*.so
+.idea/
+.vscode/
+*.swp
+*.swo
+*~

trnblas-0.2.0/CHANGELOG.md

@@ -0,0 +1,46 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.2.0] - 2026-04-11
+
+### Added
+
+- MkDocs Material documentation site at
+  [scttfrdmn.github.io/trnblas](https://scttfrdmn.github.io/trnblas/) with
+  Installation, Quickstart, API reference (Level 1/2/3, NKI backend), and
+  Architecture pages.
+- GitHub Actions CI matrix (Python 3.10, 3.11, 3.12).
+- Neuron hardware CI workflow scaffold (`workflow_dispatch`) — SSM wiring
+  deferred until a persistent CI Trainium instance is available.
+- PyPI publishing workflow (OIDC trusted publishers, sdist + wheel on
+  release) — matches trnfft pattern.
+- Benchmark suite scaffold (`benchmarks/bench_blas.py`, pytest-benchmark).
+- Issue and PR templates under `.github/`.
+- README badges — CI status, PyPI version, Python versions, License, Docs.
+- Cross-link to trnblas in trnfft's Related Projects table
+  (`scttfrdmn/trnfft@7330b3f`).
+
+## [0.1.0] - 2026-04-11
+
+### Added
+
+- Level 1 BLAS: `axpy`, `dot`, `nrm2`, `scal`, `asum`, `iamax`.
+- Level 2 BLAS: `gemv`, `symv`, `trmv`, `ger`.
+- Level 3 BLAS: `gemm`, `batched_gemm`, `symm`, `syrk`, `trsm`, `trmm`.
+- NKI dispatch layer with `auto`, `pytorch`, and `nki` backend selection.
+- NKI GEMM kernel stub with stationary tile reuse strategy (scaffolded for
+  on-hardware validation on trn1/trn2).
+- DF-MP2 example (`examples/df_mp2.py`) demonstrating the Janesko/TCU use case
+  with half-transform GEMMs, Cholesky, triangular solve, and energy evaluation.
+- Test suite covering Level 1/2/3 BLAS correctness against PyTorch/NumPy
+  references, with SPD matrix fixtures for symmetric/triangular routines.
+
+[Unreleased]: https://github.com/scttfrdmn/trnblas/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/scttfrdmn/trnblas/releases/tag/v0.2.0
+[0.1.0]: https://github.com/scttfrdmn/trnblas/commits/main

trnblas-0.2.0/CLAUDE.md

@@ -0,0 +1,105 @@
+# trnblas
+
+BLAS operations for AWS Trainium via NKI.
+Part of the trn-* scientific computing suite by Playground Logic.
+
+## What This Is
+
+A cuBLAS-equivalent for Trainium. Provides Level 1 (vector), Level 2
+(matrix-vector), and Level 3 (matrix-matrix) BLAS operations with NKI
+kernel acceleration on the Tensor Engine.
+
+**Primary use case:** DF-MP2 quantum chemistry with Prof. Ben Janesko at TCU.
+Molecules with >3000 basis functions require sustained GEMM throughput for
+tensor contractions. On 192 Zen4 cores (c8a.48xlarge), a single calculation
+takes ~24 hours at ~$200. trnblas on Trainium targets 3-4x cost reduction
+by exploiting the systolic array for the GEMM-dominated hot path.
+
+## Architecture
+
+```
+trnblas/
+├── trnblas/
+│   ├── __init__.py          # Re-exports all BLAS operations
+│   ├── level1.py            # axpy, dot, nrm2, scal, asum, iamax
+│   ├── level2.py            # gemv, symv, trmv, ger
+│   ├── level3.py            # gemm, batched_gemm, symm, syrk, trsm, trmm
+│   └── nki/
+│       ├── __init__.py
+│       └── dispatch.py      # auto/pytorch/nki dispatch + NKI GEMM kernel
+├── tests/
+│   ├── conftest.py          # Fixtures: random matrices, SPD matrices
+│   ├── test_level1.py       # Vector operation tests
+│   ├── test_level2.py       # Matrix-vector tests
+│   └── test_level3.py       # Matrix-matrix tests (GEMM, TRSM, SYRK, etc.)
+├── examples/
+│   └── df_mp2.py            # DF-MP2 energy using trnblas (Janesko use case)
+├── pyproject.toml
+├── README.md
+├── LICENSE                  # Apache 2.0
+└── CLAUDE.md                # This file
+```
+
+## DF-MP2 → BLAS Mapping
+
+The DF-MP2 algorithm maps directly to trnblas Level 3 operations:
+
+| DF-MP2 Step | Math | trnblas Call |
+|-------------|------|-------------|
+| Half-transform | (iν\|P) = C_occ^T @ (μν\|P) | `gemm(1.0, C_occ, eri, transA=True)` |
+| MO transform | (ia\|P) = C_vir^T @ (iν\|P) | `gemm(1.0, C_vir, iv_P, transA=True)` |
+| Cholesky | J = L @ L^T | `torch.linalg.cholesky` |
+| Metric solve | L^T @ X = I | `trsm(1.0, L, I, uplo="lower", trans=True)` |
+| Metric contract | B = (ia\|Q) @ J^{-1/2} | `gemm(1.0, ia_P, J_inv_half)` |
+| Energy | T_ab = B_i @ B_j^T | `gemm(1.0, B[i], B[j], transB=True)` |
+
+## NKI GEMM Strategy
+
+The NKI GEMM kernel uses stationary tile reuse on the Tensor Engine:
+
+1. Load A tile (128×128) to SBUF as stationary operand
+2. Stream B tiles through the systolic array as moving operand
+3. Accumulate partial products in PSUM
+4. One A load serves all B tiles → 2x fewer SBUF loads vs naive
+
+For DF-MP2, the MO coefficient matrix C is the natural stationary operand
+since it's reused across all auxiliary basis indices P.
+
+## Known Gaps & Design Notes
+
+- **NKI GEMM is a stub.** Falls back to `torch.matmul` until validated on
+  trn1/trn2. The kernel scaffold is in `nki/dispatch.py`.
+
+- **No FP64.** Trainium's Tensor Engine maxes out at FP32. For chemistry
+  workloads needing higher precision, double-double arithmetic (emulated FP64
+  from two FP32 values) is the path — not implemented yet.
+
+- **Level 3 dominates.** Level 1/2 are included for API completeness but
+  won't get NKI kernels initially. The Tensor Engine is wasted on vector ops.
+
+- **DF-MP2 example uses Python loops** over occupied orbital pairs. A
+  production implementation would batch these as a single batched_gemm call.
+
+## Dependencies
+
+- `torch>=2.1` — tensor operations and CPU fallback
+- `numpy>=1.24` — test reference
+- `neuronxcc` — NKI kernels (optional, only on Neuron hardware)
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+pytest tests/ -v               # CPU fallback mode
+python examples/df_mp2.py --demo   # Quick DF-MP2 demo
+```
+
+## Naming Convention
+
+Sibling repos in the trn-* suite:
+- `trnfft` — FFT + complex ops (https://github.com/scttfrdmn/trnfft)
+- `trnblas` — BLAS operations (this repo)
+- `trnsolver` — Linear solvers, eigendecomposition (planned)
+- `trnrand` — Random number generation (planned)
+
+All repos: Python/NKI, Apache 2.0, Playground Logic.

trnblas-0.2.0/LICENSE

@@ -0,0 +1,17 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   Copyright 2026 Playground Logic LLC
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

trnblas-0.2.0/PKG-INFO

@@ -0,0 +1,139 @@
+Metadata-Version: 2.4
+Name: trnblas
+Version: 0.2.0
+Summary: BLAS operations for AWS Trainium via NKI
+Author-email: Scott Friedman <scott@playgroundlogic.co>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/scttfrdmn/trnblas
+Project-URL: Repository, https://github.com/scttfrdmn/trnblas
+Project-URL: Issues, https://github.com/scttfrdmn/trnblas/issues
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.1
+Requires-Dist: numpy>=1.24
+Provides-Extra: neuron
+Requires-Dist: neuronxcc>=2.15; extra == "neuron"
+Requires-Dist: torch-neuronx>=2.1; extra == "neuron"
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0; extra == "dev"
+Requires-Dist: pytest-benchmark>=4.0; extra == "dev"
+Requires-Dist: scipy>=1.11; extra == "dev"
+Dynamic: license-file
+
+# trnblas
+
+[![CI](https://github.com/scttfrdmn/trnblas/actions/workflows/ci.yml/badge.svg)](https://github.com/scttfrdmn/trnblas/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/trnblas)](https://pypi.org/project/trnblas/)
+[![Python](https://img.shields.io/pypi/pyversions/trnblas)](https://pypi.org/project/trnblas/)
+[![License](https://img.shields.io/github/license/scttfrdmn/trnblas)](LICENSE)
+[![Docs](https://img.shields.io/badge/docs-mkdocs-blue)](https://scttfrdmn.github.io/trnblas/)
+
+BLAS operations for AWS Trainium via NKI (Neuron Kernel Interface).
+
+Trainium ships no BLAS library. `trnblas` provides Level 1-3 BLAS operations with NKI kernel acceleration on the Tensor Engine, targeting scientific computing workloads that are GEMM-dominated.
+
+Part of the **trn-\*** scientific computing suite by [Playground Logic](https://playgroundlogic.co).
+
+## Why
+
+NVIDIA has cuBLAS with 152 optimized routines. Trainium has `torch.matmul`. That's fine for ML training but insufficient for scientific computing codes that need TRSM, SYRK, SYMM, and batched GEMM with specific transpose/scaling semantics.
+
+trnblas closes this gap — same BLAS API surface, NKI-accelerated GEMM on Trainium, PyTorch fallback everywhere else.
+
+## Install
+
+```bash
+pip install trnblas
+
+# With Neuron hardware support
+pip install trnblas[neuron]
+```
+
+## Usage
+
+```python
+import torch
+import trnblas
+
+# Level 3 — Matrix multiply (the hot path)
+C = trnblas.gemm(alpha=1.0, A=A, B=B, beta=0.5, C=C_init, transA=True)
+
+# Batched GEMM (DF-MP2 tensor contractions)
+C = trnblas.batched_gemm(1.0, A_batch, B_batch)
+
+# Symmetric matrix multiply (Fock builds)
+F = trnblas.symm(1.0, density, H_core, side="left")
+
+# Triangular solve (Cholesky-based density fitting)
+X = trnblas.trsm(1.0, L, B, uplo="lower")
+
+# Symmetric rank-k update (metric construction)
+J = trnblas.syrk(1.0, integrals, trans=True)
+
+# Level 2 — Matrix-vector
+y = trnblas.gemv(1.0, A, x, beta=1.0, y=y)
+
+# Level 1 — Vector operations
+y = trnblas.axpy(alpha, x, y)
+d = trnblas.dot(x, y)
+n = trnblas.nrm2(x)
+```
+
+## DF-MP2 Example
+
+```bash
+# Run the density-fitted MP2 example (Janesko/TCU use case)
+python examples/df_mp2.py --demo
+python examples/df_mp2.py --nbasis 100 --nocc 20
+```
+
+The example demonstrates all core BLAS operations in a realistic quantum chemistry workflow: Cholesky factorization, triangular solve, half-transform GEMMs, metric contraction, and energy evaluation.
+
+## Operations
+
+| Level | Operation | Description |
+|-------|-----------|-------------|
+| 1 | `axpy` | y = αx + y |
+| 1 | `dot` | x^T y |
+| 1 | `nrm2` | ‖x‖₂ |
+| 1 | `scal` | x = αx |
+| 1 | `asum` | Σ\|xᵢ\| |
+| 1 | `iamax` | argmax \|xᵢ\| |
+| 2 | `gemv` | y = α op(A) x + βy |
+| 2 | `symv` | y = α A x + βy (A symmetric) |
+| 2 | `trmv` | x = op(A) x (A triangular) |
+| 2 | `ger` | A = α x yᵀ + A |
+| 3 | `gemm` | C = α op(A) op(B) + βC |
+| 3 | `batched_gemm` | Batched GEMM |
+| 3 | `symm` | C = α A B + βC (A symmetric) |
+| 3 | `syrk` | C = α A Aᵀ + βC |
+| 3 | `trsm` | Solve op(A) X = αB |
+| 3 | `trmm` | B = α op(A) B |
+
+## Status
+
+- [x] Level 1-3 BLAS with PyTorch backend
+- [x] GEMM with NKI dispatch stub
+- [x] DF-MP2 example (Janesko/TCU use case)
+- [ ] NKI GEMM kernel validation on trn1/trn2
+- [ ] NKI GEMM with stationary tile reuse
+- [ ] Batched GEMM NKI kernel
+- [ ] Double-double FP64 emulation
+- [ ] Benchmarks vs cuBLAS
+
+## Related Projects
+
+| Project | What |
+|---------|------|
+| [trnfft](https://github.com/scttfrdmn/trnfft) | FFT + complex ops for Trainium (Williamson/OSU use case) |
+| trnsolver *(planned)* | Linear solvers and eigendecomposition |
+
+## License
+
+Apache 2.0 — Playground Logic LLC

trnblas-0.2.0/README.md

@@ -0,0 +1,111 @@
+# trnblas
+
+[![CI](https://github.com/scttfrdmn/trnblas/actions/workflows/ci.yml/badge.svg)](https://github.com/scttfrdmn/trnblas/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/trnblas)](https://pypi.org/project/trnblas/)
+[![Python](https://img.shields.io/pypi/pyversions/trnblas)](https://pypi.org/project/trnblas/)
+[![License](https://img.shields.io/github/license/scttfrdmn/trnblas)](LICENSE)
+[![Docs](https://img.shields.io/badge/docs-mkdocs-blue)](https://scttfrdmn.github.io/trnblas/)
+
+BLAS operations for AWS Trainium via NKI (Neuron Kernel Interface).
+
+Trainium ships no BLAS library. `trnblas` provides Level 1-3 BLAS operations with NKI kernel acceleration on the Tensor Engine, targeting scientific computing workloads that are GEMM-dominated.
+
+Part of the **trn-\*** scientific computing suite by [Playground Logic](https://playgroundlogic.co).
+
+## Why
+
+NVIDIA has cuBLAS with 152 optimized routines. Trainium has `torch.matmul`. That's fine for ML training but insufficient for scientific computing codes that need TRSM, SYRK, SYMM, and batched GEMM with specific transpose/scaling semantics.
+
+trnblas closes this gap — same BLAS API surface, NKI-accelerated GEMM on Trainium, PyTorch fallback everywhere else.
+
+## Install
+
+```bash
+pip install trnblas
+
+# With Neuron hardware support
+pip install trnblas[neuron]
+```
+
+## Usage
+
+```python
+import torch
+import trnblas
+
+# Level 3 — Matrix multiply (the hot path)
+C = trnblas.gemm(alpha=1.0, A=A, B=B, beta=0.5, C=C_init, transA=True)
+
+# Batched GEMM (DF-MP2 tensor contractions)
+C = trnblas.batched_gemm(1.0, A_batch, B_batch)
+
+# Symmetric matrix multiply (Fock builds)
+F = trnblas.symm(1.0, density, H_core, side="left")
+
+# Triangular solve (Cholesky-based density fitting)
+X = trnblas.trsm(1.0, L, B, uplo="lower")
+
+# Symmetric rank-k update (metric construction)
+J = trnblas.syrk(1.0, integrals, trans=True)
+
+# Level 2 — Matrix-vector
+y = trnblas.gemv(1.0, A, x, beta=1.0, y=y)
+
+# Level 1 — Vector operations
+y = trnblas.axpy(alpha, x, y)
+d = trnblas.dot(x, y)
+n = trnblas.nrm2(x)
+```
+
+## DF-MP2 Example
+
+```bash
+# Run the density-fitted MP2 example (Janesko/TCU use case)
+python examples/df_mp2.py --demo
+python examples/df_mp2.py --nbasis 100 --nocc 20
+```
+
+The example demonstrates all core BLAS operations in a realistic quantum chemistry workflow: Cholesky factorization, triangular solve, half-transform GEMMs, metric contraction, and energy evaluation.
+
+## Operations
+
+| Level | Operation | Description |
+|-------|-----------|-------------|
+| 1 | `axpy` | y = αx + y |
+| 1 | `dot` | x^T y |
+| 1 | `nrm2` | ‖x‖₂ |
+| 1 | `scal` | x = αx |
+| 1 | `asum` | Σ\|xᵢ\| |
+| 1 | `iamax` | argmax \|xᵢ\| |
+| 2 | `gemv` | y = α op(A) x + βy |
+| 2 | `symv` | y = α A x + βy (A symmetric) |
+| 2 | `trmv` | x = op(A) x (A triangular) |
+| 2 | `ger` | A = α x yᵀ + A |
+| 3 | `gemm` | C = α op(A) op(B) + βC |
+| 3 | `batched_gemm` | Batched GEMM |
+| 3 | `symm` | C = α A B + βC (A symmetric) |
+| 3 | `syrk` | C = α A Aᵀ + βC |
+| 3 | `trsm` | Solve op(A) X = αB |
+| 3 | `trmm` | B = α op(A) B |
+
+## Status
+
+- [x] Level 1-3 BLAS with PyTorch backend
+- [x] GEMM with NKI dispatch stub
+- [x] DF-MP2 example (Janesko/TCU use case)
+- [ ] NKI GEMM kernel validation on trn1/trn2
+- [ ] NKI GEMM with stationary tile reuse
+- [ ] Batched GEMM NKI kernel
+- [ ] Double-double FP64 emulation
+- [ ] Benchmarks vs cuBLAS
+
+## Related Projects
+
+| Project | What |
+|---------|------|
+| [trnfft](https://github.com/scttfrdmn/trnfft) | FFT + complex ops for Trainium (Williamson/OSU use case) |
+| trnsolver *(planned)* | Linear solvers and eigendecomposition |
+
+## License
+
+Apache 2.0 — Playground Logic LLC