Make FASTGREEDY=1 compatible with non-greedy requests (argmax fastpath always off)

LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0 unconditionally. FASTGREEDY=1 now
means: reasoning off, deterministic sampler defaults, compatible with
all non-greedy requests (temp != 0, penalties, reasoning budget). This
is the card's 'compatible mode' profile.

scripts/frank-default-large.env

@@ -84,11 +84,10 @@ SPEC_DRAFT_P_SPLIT=0.10
 #
 # FASTGREEDY mode: set to 1 for fast-greedy benchmark (card's default profile)
 # FASTGREEDY=0: production chat config (reasoning on, argmax off)
-# FASTGREEDY=1: reasoning off, LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1, deterministic
+# FASTGREEDY=1: reasoning off, deterministic sampler, argmax fastpath off
-#   sampler. Mirrors the card's 'fast greedy (default)' benchmark config.
+#   (compatible mode). Mirrors the card's 'compatible mode' profile — the
-# 2026-08-22: flipped to 1 for real (operator decision). NOTE: the fast
+#   same as production minus reasoning. Non-greedy requests work fine.
-# greedy path rejects non-greedy requests (temp != 0, top_k > 1, penalties,
+# 2026-08-22: flipped to 1 for real (operator decision).
-# reasoning budgets) with 400. Chat clients must send temperature 0.
 FASTGREEDY=1
 
 # For fast-greedy bench: set FASTGREEDY=1 above. This flips to: reasoning off,

scripts/serve-frank-large.sh

@@ -101,16 +101,13 @@ export GGML_CUDA_GRAPH_OPT=0
 # affect the Kairic Edge IU4 lane (prompt/verification route via
 # --kairic-edge) nor the MTP draft fast path below.
 # FASTGREEDY env var (0 = production chat, 1 = fast-greedy bench from
-# card's default profile). When FASTGREEDY=1: reasoning off, argmax
+# card's default profile). When FASTGREEDY=1: reasoning off, deterministic
-# fastpath active, deterministic sampler. Compatible mode (KAIRIC_EDGE
+# sampler. The argmax fastpath is always off (compatible mode) so non-greedy
-# compatibility mode) only disables the target greedy argmax fast path;
+# requests (temp != 0, penalties, reasoning budget) work correctly.
-# reasoning is disabled separately.
+# The card calls this "compatible mode" — it disables only the target greedy
+# argmax fast path and keeps everything else (reasoning, MTP, caching).
 FASTGREEDY="${FASTGREEDY:-0}"
-if [ "$FASTGREEDY" = "1" ]; then
-    export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1
-else
 export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=0
-fi
 export LLAMA_MTP_CPU_ARGMAX_FASTPATH=1
 export PROMPTFORGE_TARGET_ONLY=0
 export PROMPTFORGE_MODE=iu4_ffn
@@ -136,9 +133,10 @@ export PROMPTFORGE_ENABLE_IU4_DECODE_M2_M5=0
 export PROMPTFORGE_OUTPUT_K8_STRICT_GREEDY=0
 
 # Sampling & reasoning defaults — production (FASTGREEDY=0) vs fast-greedy
-# (FASTGREEDY=1). Fast greedy uses the card's default deterministic sampler
+# (FASTGREEDY=1). Both use argmax fastpath off (compatible mode); the only
-# with argmax fastpath active; production uses Qwen live values with
+# difference is reasoning: off (FASTGREEDY=1) vs on/deepseek/budget-unlimited
-# reasoning budget and argmax fastpath off.
+# (FASTGREEDY=0). Sampling defaults are identical (temp 0, top_p 1, top_k 0,
+# min_p 0) and clients override per request as the card recommends.
 if [ "$FASTGREEDY" = "1" ]; then
     TEMP=0
     TOP_P=1.0