aether-models/aether-kit-phi2

Phi-2 for Aether Kit

A macOS Core AI conversion of microsoft/phi-2, at upstream revision 9a5117d15e407c48fa40f8adbac1244abae52d6f. Aether Kit source: 1881cf4.

Requirements and usage

The mac/ subdirectory is a registered Aether language bundle. It requires Apple silicon, macOS 27 or newer, and an Aether Kit checkout containing the Phi-2 spike (1881cf4 or a descendant). Published Aether Kit 0.0.4 lacks the padded-cache support required by this model. No iOS bundle is qualified.

This is a base completion model, with no chat template. Use AetherLanguageCoreAI.LanguageGenerationSession, rather than the chat facade (which reports unsupportedTemplate). No instruction tuning was added.

The context limit is 2,048 tokens, including prompt and continuation, with maximum prefill chunks of 64 tokens. Weights, KV caches and logits use fp16 storage; projections, normalization, residuals and attention use float32 arithmetic. This conversion is not int8/int4 quantized.

import AetherLanguageCoreAI
import AetherLanguageTokenizers

let tokenizer = try await LanguageTokenizer.open(bundle: bundleURL)
let session = try await LanguageGenerationSession.open(bundle: bundleURL, platform: .macos)
let result = try await session.generate(
    prompt: tokenizer.encode("def fibonacci(n):"), maximumTokens: 64)
print(tokenizer.decode(result.tokens))
await session.unload()

Download mac/ at an immutable Hub revision and use that local directory as bundleURL. A pinned catalog entry will be maintained in Spikes/phi2/catalog-macos.json. Model SHA-256: 8b2a3f6388406f183163440449aced556008f05e6b7a0308b1d423bc2ce79085.

Qualification and limits

Evidence is included under verification/; reproduction commands are in Spikes/phi2. The converter used Torch 2.11.0, Transformers 4.57.3 and coreai-torch 0.4.2. The recorded host ran macOS build 26B123, Core AI architecture h17c.

• Both float32 port contracts match all 64 reference argmax rows on each of three prompts (7, 280 and 224 tokens), with cosine >= 0.999999 and absolute tolerance 0.01. Upstream Transformers itself varies by up to 0.00815 between full-sequence and cached/chunked float32 execution on the long prompt.
• The compiled checkpoint passes the Core AI CPU logits/cache self-check: exact argmax, no lost cache rows, minimum cosine 0.99298.
• Native GPU execution matches all 192 forced-token choices and all three 64-token continuations exactly, repeatably across prefill widths 7 and 16; minimum native cosine is 0.99847. Exact agreement on these cases does not promise bit-exact float32 generation for arbitrary prompts.
• The 224-token case exercises padded-cache growth from 256 to 512 rows.
• The Swift Release regression suite reports 714 passing tests; existing GPT-2 native qualification also passes with the stride-aware cache implementation.

This experimental v0.0.6 development asset is qualified on these prompts on one Mac, not for general task quality, all Apple devices, or iOS. Phi-2 retains the capabilities and limitations described in its upstream model card.

License and provenance

Original model and tokenizer assets retain the MIT terms in LICENSE, Copyright (c) Microsoft Corporation. NOTICE identifies conversion and architecture sources. Aether Kit's conversion tooling is Apache-2.0.