Historical native CUDA path (outside the current Hailo-only mission): the Gemma 4 path uses the repository's locked Mojo 1.0.0 / MAX 26.5.0
environment (pixi.lock). MAX owns CUDA contexts, buffers and launches; all
model operations and tokenization are Mojo code. NVIDIA's system driver is a
prerequisite. No external inference engine or Python runtime is used for chat.
The loader uses Linux libc file mapping and bounded process/file operations.
Hugging Face transfers use the system curl (HTTPS certificate verification,
redirect restrictions, resume, optional parallel ranges), sha256sum (GNU
coreutils), and Linux procfs descriptor paths so transfer, verification and
publication remain bound to the opened staging inode.
Content-addressed model-store import and verification also use sha256sum,
reading the exact inherited open descriptor through Linux procfs.
Model-store gc uses GNU find (findutils) over an inherited procfs directory
descriptor with NUL-delimited output; it never interpolates a shell command.
These are separate executables invoked with argv, never a shell command string.
They are justified because this toolchain has no maintained native HTTPS client;
argv-only process execution is isolated in the loader and
core/posix_process.mojo. Curl must support --parallel
for --connections > 1 (curl 7.66+). No executable is vendored.
scripts/offline.py prepare records SHA-256 and byte size for itself,
aesir.config.json, pixi.lock, the prepared Mojo compiler, the four named MAX
runtime libraries used by the launch check, the native executable, the selected
catalog blobs, and the currently resolved Pixi executable/version. This is a
bounded prepared-checkout inventory;
it does not enumerate every ELF/system/driver transitive dependency.
Test-only tools: Python standard library for orchestration, tokenizers 0.22.2
for independently generated tokenizer fixtures, and gguf 0.19.0 with NumPy
for independent real-weight dequantization/matvec expectations. These are never
imported by production Mojo. See THIRD_PARTY_NOTICES.md for the pinned MIT
architecture/quantization reference. Public model weights are external artifacts
under their model license, pinned in the task contract and GEMMA4_CUDA.md.
The owner said “proceed” after the direct HailoRT route, compatible compiled model artifact, and first device smoke were disclosed. This accepts the narrow C ABI/C++ dependency exception recorded in DEPENDENCY_POLICY.md and D-002/D-005 in docs/HAILO10H_DECISION_LOG.md. It supersedes historical study-only/no-runtime/no-HEF restrictions for this route. The inference target remains Hailo-10H.
| Dependency / artifact | Qualified version / identity | Purpose and boundary | Origin / license |
|---|---|---|---|
| HailoRT GenAI C++ API and shared runtime | Installed 5.3.0; qualification is limited to this version | Direct in-process Hailo LLM execution through aesir_engine/hailo10h_bridge.cpp. Linked system dependency; vendor source is not vendored. The Mojo boundary carries C-compatible handles, strings, sizes, and status values. |
Hailo Technologies, official source. Installed GenAI hailo/genai/llm/llm.hpp header and hailort package copyright record identify MIT. Preserve the license/copyright notice if redistributing covered material. |
GNU C++ standard runtime (libstdc++) |
System runtime used with GCC 12 on the qualified host | Required by the bridge and HailoRT C++ ABI. C++ types and exceptions stay inside the bridge; no runtime source is vendored. | Free Software Foundation / GCC; installed libstdc++6 package copyright record identifies GPLv3 or later with GCC Runtime Library Exception 3.1. |
OpenSSL EVP digest API (libcrypto) |
Installed OpenSSL 3.0.22; slice 138 qualification in progress | Non-LLM SHA-256 model-file integrity only, bound to an opened regular HEF descriptor. The installed streaming digest API avoids implementing cryptography in the bridge or spawning a separate hashing process. Linked system library; no vendor source copied, no network operation, and no package installation for this change. | OpenSSL Project, Apache-2.0 as identified by the installed libssl3 package copyright record. Official EVP digest API; see THIRD_PARTY_NOTICES.md. |
| Hailo PCIe driver and device firmware | Installed 5.3.0 on the observed Hailo-10H | System device prerequisites managed separately from Æsir; not bundled by this integration. | Installed Hailo system stack. Package/source and redistribution terms must be retained for any future package that ships them. |
| Qwen2.5-1.5B-Instruct compiled Hailo artifact | Official Hailo 5.3.0 artifact; 2,464,582,493 bytes; SHA-256 0856a10c2b804cfe1590b80aa5e28369eb7169a9b36de836420a0b8cdf5b8d05 |
External local GenAI HEF for the narrow device smoke. Not committed, vendored, or distributed by the source change. Requires a compatible HailoRT/model pair; no GGUF CPU fallback is admitted. | Official Hailo artifact. Model and compiled-artifact redistribution terms still need a retained license record before packaging/distribution; HailoRT's MIT license does not establish the model's license. |
This profile uses the host for application orchestration and session/result ownership. HailoRT GenAI's installed header documents full LLM-pipeline device offload; that API contract is distinct from independently exposed per-layer execution telemetry. The previous eight-token smoke is narrow runtime evidence, not a portability, quality, or complete dependency-closure claim. Bounded canonical build, lifetime, timeout, admission, and prior-context restoration evidence are recorded in slices 135–137. Slice 138 file-identity-bound fingerprint/provenance changes are being qualified: the intended Linux-specific path loads HailoRT from the already opened descriptor through procfs and checks file metadata before/after model loading. This intended behavior is not yet marked accepted.