docs: refresh honesty notes for rewired demos, consolidation audit, ServingTokenizer

Co-Authored-By: Claude Fable 5 <[email protected]>
2026-07-10 03:52:13 -07:00
parent 68481ef314
commit 74d3db7ee7
1 changed files with 1 additions and 1 deletions
@@ -158,4 +158,4 @@ All in `crates/training/rtx-transformers/src/ssl/`. 163 tests passing.
 Production-ready. 113 crates. Full transformer stack, Flash Attention (v2+v3), MoE, speculative decoding (Medusa/EAGLE/Lookahead), federated learning, NAS, medical/neuroimaging/scientific domain stacks, complete JEPA platform (Batches 20–26). 30+ demo applications. 13,000+ tests.
-Known honesty notes (2026-07-09 sweep): `rtx-backend-sycl` is an experimental skeleton (ops return NotImplemented). Some demos are pure simulations that don't exercise framework crates (`rtx-distllm-demo`, `rtx-model-zoo`'s MockInferenceEngine, `rtx-inference-profiler`). MoE and flash-attention have duplicated implementations pending consolidation (see `docs/consolidation.md`). Serving/streaming return 503/error until an engine+model is attached (no more mock responses); tokenization at the serving boundary is byte-level until a real tokenizer is threaded through.
+Known honesty notes (2026-07-09 sweep, updated 2026-07-10): `rtx-backend-sycl` is an experimental skeleton (ops return NotImplemented). The formerly-simulated demos now run real compute (`rtx-distllm-demo` real tensors + measured metrics with only the network topology simulated; `rtx-model-zoo` drives a real InferenceEngine with untrained weights and labeled toy output proxies; `rtx-inference-profiler` profiles real matmul/softmax). MoE/flash-attention duplication was audited — see `docs/consolidation.md`: the flagged sites were scaffolding/stubs, not duplicates; `revolutionary/orchestrator_core.rs` execute paths are no-op stubs and `tensor_core_kernels.rs` is a planner with no kernels. Serving/streaming return 503/error until an engine+model is attached; tokenization uses `ServingTokenizer` (HuggingFace tokenizer.json or byte-level fallback).
@@ -158,4 +158,4 @@ All in `crates/training/rtx-transformers/src/ssl/`. 163 tests passing.

	`Production-ready. 113 crates. Full transformer stack, Flash Attention (v2+v3), MoE, speculative decoding (Medusa/EAGLE/Lookahead), federated learning, NAS, medical/neuroimaging/scientific domain stacks, complete JEPA platform (Batches 20–26). 30+ demo applications. 13,000+ tests.`	`Production-ready. 113 crates. Full transformer stack, Flash Attention (v2+v3), MoE, speculative decoding (Medusa/EAGLE/Lookahead), federated learning, NAS, medical/neuroimaging/scientific domain stacks, complete JEPA platform (Batches 20–26). 30+ demo applications. 13,000+ tests.`

	Known honesty notes (2026-07-09 sweep): `rtx-backend-sycl` is an experimental skeleton (ops return NotImplemented). Some demos are pure simulations that don't exercise framework crates (`rtx-distllm-demo`, `rtx-model-zoo`'s MockInferenceEngine, `rtx-inference-profiler`). MoE and flash-attention have duplicated implementations pending consolidation (see `docs/consolidation.md`). Serving/streaming return 503/error until an engine+model is attached (no more mock responses); tokenization at the serving boundary is byte-level until a real tokenizer is threaded through.	Known honesty notes (2026-07-09 sweep, updated 2026-07-10): `rtx-backend-sycl` is an experimental skeleton (ops return NotImplemented). The formerly-simulated demos now run real compute (`rtx-distllm-demo` real tensors + measured metrics with only the network topology simulated; `rtx-model-zoo` drives a real InferenceEngine with untrained weights and labeled toy output proxies; `rtx-inference-profiler` profiles real matmul/softmax). MoE/flash-attention duplication was audited — see `docs/consolidation.md`: the flagged sites were scaffolding/stubs, not duplicates; `revolutionary/orchestrator_core.rs` execute paths are no-op stubs and `tensor_core_kernels.rs` is a planner with no kernels. Serving/streaming return 503/error until an engine+model is attached; tokenization uses `ServingTokenizer` (HuggingFace tokenizer.json or byte-level fallback).