HiramKHackenbacker commited on
Commit
cbf6fbc
·
verified ·
1 Parent(s): 668d018

docs: stable/ folders in Status, oracle verification line, known-issue points at available fix

Browse files
Files changed (1) hide show
  1. README.md +11 -2
README.md CHANGED
@@ -37,8 +37,9 @@ produces a bare input-arity error naming `ple_table`/`ple_scale`.
37
  > because previously seen lengths replay from cache at no cost. Long free-form
38
  > generations grow without bound until the process exits; no in-process mitigation
39
  > exists (token caps only postpone it, and shape-length reuse is the only stable state).
40
- > A re-export onto a shape-stable contract (fixed-capacity KV read with an explicit
41
- > attention mask) is planned and will replace these bundles' decode graphs.
 
42
 
43
  ## Provenance
44
 
@@ -212,6 +213,7 @@ reachable. 15k was not attempted.
212
  | `ctx16384/…_tbl_pf64` | **QUALIFIED BY EQUIVALENCE** — as above; free-form at this depth was not attempted. |
213
  | `w4a8/gemma4_e4b_qat_decode_int4lin_a8_tbl_pf64` | **EXPERIMENTAL** — built, unmeasured. Gate: a Mac-side oracle/parity check plus a device benchmark. |
214
  | `gpu-pipelined/gemma4_e4b_qat_decode_int4lin` | **BLOCKED** — does not load on any current code path. |
 
215
 
216
  ### About the `w4a8` bundle
217
 
@@ -292,3 +294,10 @@ trade for the unbounded growth it replaces.
292
 
293
  Unlike the `ctx*/` folders, the context ceiling here is **encoded in the graph**: a
294
  different window needs a different export, not a manifest edit.
 
 
 
 
 
 
 
 
37
  > because previously seen lengths replay from cache at no cost. Long free-form
38
  > generations grow without bound until the process exits; no in-process mitigation
39
  > exists (token caps only postpone it, and shape-length reuse is the only stable state).
40
+ > A re-export onto a shape-stable contract that removes this growth is available in this
41
+ > repository under `stable/` see its section below. It requires updated host support
42
+ > (in progress); on current hosts, use the bundles above for guided generation only.
43
 
44
  ## Provenance
45
 
 
213
  | `ctx16384/…_tbl_pf64` | **QUALIFIED BY EQUIVALENCE** — as above; free-form at this depth was not attempted. |
214
  | `w4a8/gemma4_e4b_qat_decode_int4lin_a8_tbl_pf64` | **EXPERIMENTAL** — built, unmeasured. Gate: a Mac-side oracle/parity check plus a device benchmark. |
215
  | `gpu-pipelined/gemma4_e4b_qat_decode_int4lin` | **BLOCKED** — does not load on any current code path. |
216
+ | `stable/…_tbl_pf64_stable_c16384` | **EXPERIMENTAL** — shape-stable decode contract; removes the per-token memory growth by design; requires host support for absolute-position inputs (in progress); no on-device measurements yet. |
217
 
218
  ### About the `w4a8` bundle
219
 
 
294
 
295
  Unlike the `ctx*/` folders, the context ceiling here is **encoded in the graph**: a
296
  different window needs a different export, not a manifest edit.
297
+
298
+ **Verification.** Before conversion, the stable graph was checked against the shipped
299
+ graph in eager torch on the same weights: greedy argmax identical on all tested decode
300
+ steps (including positions past the 512-token sliding window), worst logits cosine
301
+ 0.99999 (fp16), and the prefill path bit-identical in fp32. On-device behavior is
302
+ unmeasured until host support lands.
303
+