You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 082a625
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: docs/OPENVINO_UPSTREAM_GAPS_AND_PR_ROADMAP.md
+5-4Lines changed: 5 additions & 4 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -28,7 +28,7 @@ This roadmap consolidates all 14 gaps into a single engineering authority docume
28
28
|**GAP-01**|`openvino::op`|**Critical**| Missing native `RMSNorm` in standard opsets |`src/core/include/openvino/op/`| PR-01: Native `RMSNorm` op with fused GPU/NPU kernels |
29
29
|**GAP-02**|`xml_util`|**High**| Raw `i < m_inputs.size()` index out-of-range crash |`src/core/xml_util/src/xml_deserialize_util.cpp`|**[RESOLVED & PR SUBMITTED]**[PR #39429](https://github.com/openvinotoolkit/openvino/pull/39429): Actionable disconnected port diagnostics |
|**GAP-04**|`openvino.genai`|**High**| Aborts with `invalid vector subscript` on invalid IR |`openvino_genai/src/llm_pipeline.cpp`|PR-04: DAG topological pre-flight validation |
31
+
|**GAP-04**|`openvino.genai`|**High**| Aborts with `invalid vector subscript` on invalid IR |`openvino_genai/src/llm_pipeline.cpp`|**[RESOLVED & PR SUBMITTED]**[PR #4807](https://github.com/openvinotoolkit/openvino.genai/pull/4807): DAG topological pre-flight validation |
32
32
|**GAP-05**|`openvino.genai`|**Critical**| Zero native GGUF container ingestion |`openvino_genai/src/`| PR-05: Zero-dependency C++ GGUF container reader |
@@ -171,16 +171,17 @@ This roadmap consolidates all 14 gaps into a single engineering authority docume
171
171
- Validated 100% pass on negative diagnostic tests (undersized FP32 buffer, safe unnamed fallbacks, sub-byte U4 bitwidth ceiling delta) and 0 regressions on full-scale production models (`qwen-2.5-1.5b` with 2,059 constants and `deepseek-r1-1.5b` with 1,973 constants).
172
172
173
173
### PR-04: DAG Topological Pre-flight Validation in OpenVINO GenAI
1. Implemented `ov::genai::utils::validate_llm_model_dag` with Gate 0 (Null guard), Gate A (Role-based parameter whitelist for PRIMARY_LLM and DRAFT/EAGLE3 models), Gate B (Fuzzy logit name heuristics), and Gate C (Dual-root backward BFS from Results and Sinks).
182
183
2. Integrated defensive pre-flight hook in `StatefulLLMPipeline` and `StaticLLMPipeline` before device compilation, preventing uninformative `invalid vector subscript` aborts.
183
-
3. Added comprehensive GTest suite `tests/cpp/test_llm_pipeline_preflight.cpp` auto-discovered by CMake.
184
+
3. Added comprehensive GTest suite `tests/cpp/test_llm_pipeline_preflight.cpp` (16 test cases, 100% pass) auto-discovered by CMake.
184
185
4. Executed physical silicon verification harness `scratch/verify_gap04_dag_preflight.py` on Intel Core Ultra 5 125H silicon (Arc Xe-LPG iGPU & AI Boost NPU 3720), proving 0 regressions on `qwen-2.5-1.5b`, `deepseek-r1-1.5b`, and `qwen-0.5b-npu-certified`.
185
186
186
187
### PR-05: Native C++ Zero-Dependency GGUF Container Reader
Malformed or corrupt OpenVINO IR models (e.g. models missing `input_ids`, missing vocabulary `logits`, or with topologically disconnected parameters from aggressive pruning/quantization tools) currently bypass `ov::Core::read_model` and fail deep inside pipeline instantiation with cryptic internal runtime errors such as `std::out_of_range: invalid vector subscript` or null pointer dereferences.
3
+
4
+
This PR introduces topological pre-flight validation `ov::genai::utils::validate_llm_model_dag` in `src/cpp/src/utils.hpp` and `src/cpp/src/utils.cpp`, called during `StatefulLLMPipeline` construction (in both stateful and static pipeline paths) before KV-cache initialization and plugin compilation:
5
+
6
+
1.**Gate 0 (Null & Zero-Element Guard)**:
7
+
Asserts model pointer is non-null and contains at least one parameter and at least one result or sink.
8
+
2.**Gate A (Role-Aware Input Validation)**:
9
+
Verifies primary LLMs possess `input_ids`. For speculative draft models (`ModelRole::DRAFT_MODEL` / `EAGLE3_DRAFT`), whitelists `input_ids`, `hidden_states`, or `inputs_embeds`.
10
+
3.**Gate B (Heuristic Logits Output Verification)**:
11
+
Inspects output friendly names and tensor names for `logits`, `lm_logits`, rank $\ge 2$ single-result fallback, and draft heads.
12
+
4.**Gate C (Dual-Root Backward BFS Traversal)**:
13
+
Traverses backward from both `ov::Result` nodes and `ov::Sink` nodes (`ov::op::v6::Assign` in stateful KV-cache loops) to confirm that critical parameters are topologically connected to outputs, preventing disconnected-subgraph crashes.
14
+
15
+
When validation fails, an informative `ov::Exception` is raised detailing the model name, exact missing tensor, list of existing tensors, and actionable advice to re-export or fix the graph.
-[x] This PR follows [GenAI Contributing guidelines](https://github.com/openvinotoolkit/openvino.genai?tab=contributing-ov-file#contributing).
21
+
-[x] Tests have been updated or added to cover the new code: Added 16 GTest test cases in `tests/cpp/test_llm_pipeline_preflight.cpp` covering null models, empty models, missing inputs, missing logits, disconnected parameters, draft models, stateful sink loops, thread safety, and latency budget (<5ms).
22
+
-[x] This PR fully addresses the ticket.
23
+
-[x] I have made corresponding changes to the documentation (N/A - internal pre-flight validation without public API contract changes).
24
+
25
+
## AI Usage Disclosure:
26
+
AI assistance used: yes
27
+
If yes: Assisted in drafting initial boilerplate and exhaustive test permutations for GTest.
28
+
Human validation performed: Complete manual architectural review, MSVC 2026 build verification, 16/16 GTest suite passing, regression suite passing, and physical silicon validation on Intel Meteor Lake Core Ultra 5 125H (Arc GPU & NPU) with Optimum-Intel models (Qwen2.5-1.5B, DeepSeek-R1-1.5B, Qwen-0.5B NPU).
0 commit comments