Unify proposer to z-lab across ALL inference/eval entry points (CUDA gate + Mac harnesses default + run examples); training script + orchestration keep kakeya-baseline (how f_theta v5 was historically trained), documented in docs/k3-gpu-beta.md

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: cursoragent

docs/k3-gpu-beta.md

@@ -91,6 +91,13 @@ sliding-layer restored K/V are window-masked during decode. Both incremental
 decode and fused spec-decode measure **recall 1.0** with z-lab. (If pure
 sliding-layer restoration is ever needed, retrain f_θ on z-lab K/V.)
 
+All **inference/eval** entry points default to z-lab (`k3_e2e_gpu_bench`,
+`k3_specdecode_gpu_bench`, `k3_integrated_niah_eval`(+`_mac`),
+`k3_dflash_specdecode_eval`(+`_mac`); the gRPC server takes an explicit
+`--drafter-id`). The **f_θ training** script (`k3_f_theta_train.py`) and its
+orchestration `.sh` keep `models/dflash-kakeya-baseline` because that is how the
+shipped v5 checkpoint was historically trained.
+
 ## Notes / scope
 
 * Drafting conditions on the restored verifier hidden for committed decode tokens

scripts/research/k3_dflash_specdecode_eval_mac.py

@@ -146,7 +146,7 @@ def main() -> int:
         help="Local MLX 4-bit verifier directory (default: standard Mac path).",
     )
     ap.add_argument(
-        "--drafter-id", default="models/dflash-kakeya-baseline",
+        "--drafter-id", default="z-lab/gemma-4-26B-A4B-it-DFlash",
         help="DFlash drafter source — local path or HF id. Default: the "
              "alignment-trained baseline on main (post PR #93 + #99 merge).",
     )

scripts/research/k3_integrated_niah_eval.py

@@ -48,7 +48,7 @@
   HF_TOKEN=hf_xxx PYTHONPATH=.:sdks/python python3 \\
       scripts/research/k3_integrated_niah_eval.py \\
       --verifier-id google/gemma-4-26B-A4B-it \\
-      --drafter-id models/dflash-kakeya-baseline \\
+      --drafter-id z-lab/gemma-4-26B-A4B-it-DFlash \\
       --f-theta-dir results/research/f_theta_v1 \\
       --n-samples 10 --haystack-min-lines 60 --haystack-max-lines 80 \\
       --sink-size 4 --window-size 64 \\
@@ -91,7 +91,7 @@
 def parse_args() -> argparse.Namespace:
     ap = argparse.ArgumentParser(description=__doc__)
     ap.add_argument("--verifier-id", default="google/gemma-4-26B-A4B-it")
-    ap.add_argument("--drafter-id", default="models/dflash-kakeya-baseline")
+    ap.add_argument("--drafter-id", default="z-lab/gemma-4-26B-A4B-it-DFlash")
     ap.add_argument("--f-theta-dir", required=True,
                     help="Directory containing f_theta_config.json + f_theta_weights.pt")
     ap.add_argument("--n-samples", type=int, default=10)

scripts/research/k3_integrated_niah_eval_mac.py

@@ -19,7 +19,7 @@
 
     PYTHONPATH=.:sdks/python python3 scripts/research/k3_integrated_niah_eval_mac.py \\
         --verifier-path models/gemma-4-26B-A4B-it-mlx-4bit \\
-        --drafter-id    models/dflash-kakeya-baseline \\
+        --drafter-id    z-lab/gemma-4-26B-A4B-it-DFlash \\
         --f-theta-dir   results/research/f_theta_v5_s5_sliding \\
         --s5-exact-full-attn \\
         --n-samples 10 --haystack-min-lines 238 --haystack-max-lines 322 \\
@@ -55,7 +55,7 @@
 def parse_args() -> argparse.Namespace:
     ap = argparse.ArgumentParser(description=__doc__)
     ap.add_argument("--verifier-path", default="models/gemma-4-26B-A4B-it-mlx-4bit")
-    ap.add_argument("--drafter-id", default="models/dflash-kakeya-baseline")
+    ap.add_argument("--drafter-id", default="z-lab/gemma-4-26B-A4B-it-DFlash")
     ap.add_argument("--f-theta-dir", required=True)
     ap.add_argument("--n-samples", type=int, default=10)
     ap.add_argument("--haystack-min-lines", type=int, default=238)