liangsu9988 commited on
Commit
12b8c99
·
verified ·
1 Parent(s): 285871e

Record v2 end-to-end verdict and structural-split downgrade in RESULTS

Browse files
Files changed (1) hide show
  1. README.md +25 -24
README.md CHANGED
@@ -1,32 +1,33 @@
1
- # flashrt-flex-attention-train
2
 
3
- FlexAttention replacement training package for PI-style dual-expert
4
- transformers.
 
 
5
 
6
- Hub repo: `flashrt/flashrt-flex-attention-train`
7
 
8
- ## Public API
 
 
 
 
9
 
10
- - `flex_attention`
11
- - `flex_attention_forward`
12
- - `reference_flex_attention`
13
- - `build_block_sparse_bool_masks`
14
- - `backend_marker`
15
 
16
- ## Scope
 
 
 
 
17
 
18
- This package locks the public Tensor API and correctness harness for a native
19
- replacement of the PI052 FlexAttention/SDPA attention path:
20
 
21
- - prefix self-attention rows
22
- - action-to-prefix rows plus block-diagonal action rows
23
- - `head_dim=256`
24
- - BF16 forward/backward through PyTorch autograd fallback
25
- - detached-prefix semantics for action rows reading prefix K/V
26
- - prefix mask, prefix padding mask, action block mask, and action padding mask
27
- - automatic SDPA fallback for unsupported shapes
28
 
29
- The current implementation is the SDPA-backed training reference. It is meant
30
- to be the stable integration target for native CUDA fwd/bwd kernels; no native
31
- performance claim is made until the benchmark gates in `VALIDATION.md` pass on
32
- both A100 and RTX 5090.
 
 
1
+ # pi052 Training Optimization Workspace
2
 
3
+ This workspace records the pi052 training optimization plan, baseline tests,
4
+ profiles, and environment notes. It is intentionally separate from the
5
+ `lerobot/` source tree so optimization experiments do not mix with upstream
6
+ code changes.
7
 
8
+ Current policy:
9
 
10
+ - Main test machine: RTX 5080 server at `internal.mindon.tech:20001`.
11
+ - Thor server is parked for now; it is not the training test target.
12
+ - Hugging Face Jobs are a final validation path only, and only A100 is planned.
13
+ - HF Jobs budget target: keep early validation under a few dozen dollars.
14
+ - Prefer China-friendly mirrors and ModelScope for downloads.
15
 
16
+ Directory layout:
 
 
 
 
17
 
18
+ - `docs/`: plans, environment records, and test protocols.
19
+ - `scripts/`: helper scripts for remote setup, probing, and later test runs.
20
+ - `results/`: JSON/CSV/text outputs from tests.
21
+ - `profiles/`: torch profiler traces and summaries.
22
+ - `env/`: requirements and environment lock notes.
23
 
24
+ No pi052 model code should be edited from this workspace directly. Source
25
+ changes should happen in `../lerobot/` after the baseline tests are in place.
26
 
27
+ Primary planning documents:
 
 
 
 
 
 
28
 
29
+ - `docs/plan.md`: phase-level strategy.
30
+ - `docs/proof_tests.md`: exact tests and proof thresholds before coding.
31
+ - `docs/baseline_tests.md`: shorter baseline checklist.
32
+ - `docs/status.md`: what has already been prepared.
33
+ - `docs/proof_results_initial.md`: first synthetic proof-test results.