Repository navigation
Pin FastFlowLM at the Gemma 4 engine merge and ship the installed override. - #133
Merged
Merged
Conversation
jtuyls
force-pushed
the
users/jtuyls/flm-gemma4-release
branch
6 times, most recently
from
October 16, 2026 16:58
35d0277 to
efd339b
Compare
There was a problem hiding this comment.
🟡 Changes recommended
Shared slice dirty flags can leave data unsynchronized, and the Windows wheel launcher overrides caller PATH precedence.
2 open findings
What changed in this PR
Pins FastFlowLM to the Gemma 4 engine merge and packages sources and build support for installed engine overrides.
Changes:
- Ships Gemma 4 engine sources, link libraries, and IRON override files.
- Adds Windows compiler and engine-loading patches plus HRX sub-buffer support.
- Updates binary licensing, packaging checks, and tests.
| File | Description |
|---|---|
| tools/fastflow-sdk-qualification/licenses/FastFlowLM-TERMS.md | Removes superseded binary terms. |
| tools/fastflow-sdk-qualification/licenses/FastFlowLM-LICENSE_BINARY.txt | Adds replacement binary license. |
| tests/windows/test_fastflow_sdk_wheel_native.py | Updates license fixture. |
| tests/wheels/test_fastflow_sdk_wheel_common.py | Checks engine payload partitioning. |
| tests/wheels/test_fastflow_sdk_ops_wheels.py | Checks packaged override files. |
| tests/qualification/test_qualification_wheel.py | Verifies replacement license packaging. |
| tests/fastflowlm/test_gemma4e_engine_sources.py | Tests engine source and library staging. |
| tests/fastflowlm/test_engine_delay_load.py | Tests Windows engine loading. |
| patches/README.md | Documents new FastFlowLM patches. |
| patches/FastFlowLM/0006-engine-dll-from-path.patch | Adds PATH-first engine loading. |
| patches/FastFlowLM/0005-hrx-subbuffer-offset.patch | Adds offset-aware HRX buffer bindings. |
| patches/FastFlowLM/0004-windows-clang-cl-avx512.patch | Adjusts Windows compiler and linking options. |
| manifests/versions.toml | Updates FastFlowLM pin and notes. |
| cmake/gemma4e-override/CMakeLists.txt | Supports installed engine builds against HRX. |
| build_tools/verify_runtime_archive_hrx_cmake.py | Checks installed engine SDK contents. |
| build_tools/runtime_wheel_staging.py | Packages engine sources and exposes paths. |
| build_tools/native_dependencies.py | Separates normal and delay-load imports. |
| build_tools/fastflow_sdk_qualification_wheel.py | Stages replacement binary license. |
| build_tools/fastflow_sdk_ops_wheels.py | Ships selected Gemma override files. |
| build_tools/fastflow_sdk_archive.py | Stages and validates engine SDK payloads. |
| .gitmodules | Ignores dirty LLVM submodule contents. |
🧠 Review effort: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
jtuyls
force-pushed
the
users/jtuyls/flm-gemma4-release
branch
6 times, most recently
from
October 19, 2026 09:02
9b893d4 to
9cdacb1
Compare
jtuyls
force-pushed
the
users/jtuyls/flm-gemma4-release
branch
from
October 19, 2026 09:41
9cdacb1 to
95532b3
Compare
The product flm stays on the prebuilt engine. Upstream main stamps FLM_VERSION 1.0.9 and links minicpm_v_4_7_npu, which Windows delay-loads with the other model engines. The archive and FastFlowLM wheel carry the Gemma 4 engine sources and q4nx/mha libraries, and the ops wheel ships the IRON override files. q4_npu_eXpress and llama_npu stay normal imports. An installed q4 or mha library is left in place. Upstream replaced TERMS.md with LICENSE_BINARY.txt, so the archive, the FastFlowLM wheel, and the qualification snapshot carry that file. The qualification profiles hash the pinned model_list.json. A sub-buffer shares its parent allocation, and flush and invalidate cover only the slice that is bound. A failed invalidate records that range and continues, and a slice past its parent is rejected. Runtime::get() constructs the executable cache and the coherence map before Runtime, so process exit does not read freed statics. Fast qualification dispatches one shipped dequant or mm xclbin through those slices and checks that the bytes outside each slice stay unchanged and the output matches the recipe. The Windows wheel launcher appends its bin directories after the caller's PATH.
jtuyls
force-pushed
the
users/jtuyls/flm-gemma4-release
branch
from
October 19, 2026 09:54
95532b3 to
c409f6e
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.