Mac config: the fp16 decode (STRATA_METAL_F16=1) is the default of a resident IQ3_S or Q2_0 pack; the Coder joins --spec 3 --spec-min-p 0.5

Coder: expert-pruned 256-expert Coder, M5 Max, 5 interleaved pairs, identical greedy tokens in all 40 runs per class, over --spec 2: story -2.5%
(0/5), code +7.1%, JSON +9.3%, agent turn +12.5% (5/5 each). It has no Mac setup entry yet; the name is what a config for it would pass.

fp16 decode (PR 4807): setup writes STRATA_METAL_F16=1 into the config's env for a resident IQ3_S or Q2_0 pack (STRATA_METAL_F16=0 there turns
it off; streaming and the other sets are untouched). The owner's rule was an option unless it passes the 61-task quality suite: off / on / off
again on IQ3_S 54/61 each, on Q2_0 54/61 each, the same 7 tasks failing in all arms, no family drop, no new hallucination flag, 0 pass/fail
changes between two runs of the same arm; decode +14% (IQ3_S) and +15% (Q2_0) in the suite runs, top-1 96-97% and mean KL 0.009-0.016 against fp32
(phaseA-tests/mac-q/RESULTS.md sections 20, 22, 30).

docs/MAC.md

@@ -130,8 +130,9 @@ The ready-made Mac engine is in the app and in the Terminal's engine component,
 ## Defaults and options
 
 The config uses the Mac defaults: `--spec 2` with MTP (best on mixed text; an explicit `--spec N` in the config's `args` is kept),
-or `--spec 3 --spec-min-p 0.5` for the IQ3_S and Q2_0 sets (M5 Max, 5 interleaved pairs, same tokens, over `--spec 2`: IQ3_S story -0.9%,
-code +2.5%, JSON +16.7%, agent turn +4.9%; Q2_0 story +0.4%, code +11.5%, JSON +6.3%, agent turn +1.9%; the other sets are not measured),
+or `--spec 3 --spec-min-p 0.5` for the IQ3_S, Q2_0 and Coder sets (M5 Max, 5 interleaved pairs, same tokens, over `--spec 2`: IQ3_S story -0.9%,
+code +2.5%, JSON +16.7%, agent turn +4.9%; Q2_0 story +0.4%, code +11.5%, JSON +6.3%, agent turn +1.9%; Coder story -2.5%, code +7.1%,
+JSON +9.3%, agent turn +12.5%; the other sets are not measured),
 and the Q8_0 hyper-connection read. `STRATA_HC_Q8=0` in the config's `env` gives the exact BF16 read. The Mac engine is
 deterministic: two runs give identical numbers.
 
@@ -263,7 +264,7 @@ For testing. All are off unless set; the defaults are described above.
 | `STRATA_METAL_WIRE_ARENA=0` | Streaming, 64 GB or more: leave the filled arena pageable (default: wired) |
 | `STRATA_METAL_IGNORE_REPACK=1` | Start although a model repack or setup's preparation is running (the engine refuses by default) |
 | `STRATA_METAL_KEEP_HOST_WEIGHTS=1` | Streaming: keep the host code's own copies of the weights (2 to 3 GB) |
-| `STRATA_METAL_F16=1` | Decode windows of a resident model: the GEMVs of Q2_0, Q3_K-Q6_K and IQ4_XS and the Q2_0 routed experts in fp16. Measured on an M1 Max only: a 2-token window's GPU time 56-60 -> 40-41 ms, mean KL 0.009 against fp32 |
+| `STRATA_METAL_F16=1` / `=0` | Decode windows of a resident model: the GEMVs of Q2_0, Q3_K-Q6_K and IQ4_XS and the Q2_0 routed experts in fp16. On in the config setup writes for a resident IQ3_S or Q2_0 pack (M5 Max: decode +9 to +15%, top-1 96-97% and mean KL 0.009-0.016 against fp32; the 61-task quality suite passes the same tasks as with it off); `=0` in the config's `env` turns it off. Other sets: off unless set. M1 Max: a 2-token window's GPU time 56-60 -> 40-41 ms, mean KL 0.009 against fp32 |
 | `STRATA_METAL_TRACE=1`, `STRATA_METAL_PROFILE=1`, `STRATA_METAL_KPROF=T` | Per-window traces, a time breakdown, per-kernel GPU times |
 
 ## How it works

metal/setup_mac.py

@@ -544,7 +544,8 @@ def main(argv=None) -> int:
     S.step(7, "writing the start script")
     port = a.port or 8080
     cfg = mac_config(exe, engine_args(pack, shards, rt, ctx, stream, info["ram_gb"], model), pack, f"{fam['name']}-{model.lower()}", tag, port)
-    cfg["env"] = {"STRATA_METAL_STREAM": "1" if stream else "0"}   # the engine follows setup's measured choice
+    cfg["env"] = {"STRATA_METAL_STREAM": "1" if stream else "0",   # the engine follows setup's measured choice
+                  **_fit.engine_env(model, stream)}                # STRATA_METAL_F16=1 for a resident IQ3_S / Q2_0 (mac_fit.F16_MODELS)
     if vision == "yes":
         cfg["args"] += ["--vision"]
         cfg["vision"] = {"exe": str(vexe), "mmproj": str(mmproj), "model": str(files[0]), "gpu": vision_gpu,

strata_control/mac_fit.py

@@ -47,8 +47,23 @@ STREAM_SPEC = 3
 # A resident IQ3_S or Q2_0 pack runs a 3-token window with --spec-min-p 0.5 (the draft layer extends the window only when sure):
 # M5 Max, MTP, 5 interleaved pairs, identical tokens, over --spec 2: IQ3_S story / code / JSON / agent turn -0.9% / +2.5% / +16.7% /
 # +4.9%; Q2_0 +0.4% / +11.5% / +6.3% / +1.9% (phaseA-tests/mac-q/RESULTS.md sections 14 and 15).  Other sets are not measured: --spec 2.
-SPEC3_MODELS = ("IQ3_S", "Q2_0")
+# The expert-pruned Coder (256 experts, IQ1_M) story / code / JSON / agent turn -2.5% / +7.1% / +9.3% / +12.5%, 5 pairs, identical tokens
+# (section 31.1 of the same file; it has no Mac setup entry yet, the name is what a config for it would pass).
+SPEC3_MODELS = ("IQ3_S", "Q2_0", "CODER")
 SPEC3_MIN_P = "0.5"
+# STRATA_METAL_F16=1 (the fp16 decode windows, PR 4807) is the default of a resident IQ3_S or Q2_0 pack.  M5 Max: decode +9 to +15% (Q2_0)
+# and +9 to +12% (IQ3_S); top-1 agreement against the fp32 path 97.1% (Q2_0) / 96.2% (IQ3_S), mean KL 0.009 / 0.016 (sections 20, 22);
+# the 61-task quality suite (section 30): same tasks passed with it on and off, no family drop, no new hallucination flag, 0 pass/fail
+# changes between two runs of the same arm.  The owner's rule was an option unless it passes that suite.  STRATA_METAL_F16=0 in the config's
+# env turns it off.  Other sets (and a streaming pack, where the windows are not fp16) are not set.
+F16_MODELS = ("IQ3_S", "Q2_0")
+
+
+def engine_env(model: str, stream: bool) -> dict[str, str]:
+    """The engine variables a Mac config carries besides STRATA_METAL_STREAM: STRATA_METAL_F16=1 for a resident IQ3_S or Q2_0 pack."""
+    if not stream and (model or "").upper() in F16_MODELS:
+        return {"STRATA_METAL_F16": "1"}
+    return {}
 
 
 def spec_args(model: str, stream: bool, mtp: bool) -> list[str]:

tools/test_setup_mac.py

@@ -129,6 +129,9 @@ class Config(unittest.TestCase):
         q2 = M.engine_args(Path("/p/packs/q2_0"), shards, Path("/d/mtp/rt"), 32768, model="Q2_0")
         self.assertEqual(q2[q2.index("--spec") + 1], "3")
         self.assertEqual(q2[q2.index("--spec-min-p") + 1], "0.5")
+        coder = M.engine_args(Path("/p/packs/coder"), shards, Path("/d/mtp/rt"), 32768, model="Coder")   # 5 pairs: story -2.5%, code +7.1%, JSON +9.3%, agent +12.5%
+        self.assertEqual(coder[coder.index("--spec") + 1], "3")
+        self.assertEqual(coder[coder.index("--spec-min-p") + 1], "0.5")
         other = M.engine_args(Path("/p/packs/iq3_xxs"), shards, Path("/d/mtp/rt"), 32768, model="IQ3_XXS")   # not measured: the old default
         self.assertEqual(other[other.index("--spec") + 1], "2")
         self.assertNotIn("--spec-min-p", other)
@@ -143,6 +146,16 @@ class Config(unittest.TestCase):
         self.assertEqual(big[big.index("--spec") + 1], "3")
         self.assertNotIn("--spec-min-p", big)
 
+    def test_fp16_decode_is_set_for_a_resident_iq3_s_or_q2_0_only(self):
+        # the 61-task suite (phaseA-tests/mac-q/RESULTS.md section 30) passed the same tasks with STRATA_METAL_F16=1 as with =0
+        self.assertEqual(M._fit.engine_env("IQ3_S", False), {"STRATA_METAL_F16": "1"})
+        self.assertEqual(M._fit.engine_env("Q2_0", False), {"STRATA_METAL_F16": "1"})
+        self.assertEqual(M._fit.engine_env("iq3_s", False), {"STRATA_METAL_F16": "1"})
+        self.assertEqual(M._fit.engine_env("IQ3_S", True), {})            # streaming: the windows are not fp16
+        self.assertEqual(M._fit.engine_env("UD-Q4_K_XL", False), {})      # not measured
+        self.assertEqual(M._fit.engine_env("IQ3_XXS", False), {})
+        self.assertEqual(M._fit.engine_env("", False), {})
+
     def test_config_is_a_model_config_with_the_metal_backend(self):
         cfg = M.mac_config(Path("/r/build-metal/strata"), ["--pack", "x"], Path("/p"), "qwen3.8-flash-next-q2_0",
                            "Q2_0", 8080)