frank-small: revert to bf16 KV — the only dtype that boots Gemma4

int8_per_token_head: page-size unification fails (scale tensor breaks
the 256/512 head-dim ratio). turboquant_k8v4: backend validation fails
(no mm-prefix attention, which Gemma4's vision/audio path requires).
fp8: hardware-dead on sm_86. bf16 it is; history documented in
frank-small.env, seam pins auto + no backend pin.

scripts/frank-small.env

@@ -84,17 +84,18 @@ LISTEN_PORT=8000
 # 16k prefill chunk, 4-token MTP draft (the assistant carries 4 layers).
 MAX_MODEL_LEN=262144
 GPU_MEM_UTIL=0.95
-KV_CACHE_DTYPE=turboquant_k8v4
+KV_CACHE_DTYPE=auto
-# ATTENTION_BACKEND: Gemma4's config hook force-selects TRITON_ATTN for the
+# KV quant history (measured 2026-09-19, all on this image): bf16 (auto)
-# mixed 256/512 head dims whenever no backend is selected, and TRITON_ATTN
+# is the ONLY dtype that boots Gemma4 here. int8_per_token_head dies in
-# rejects the turboquant KV dtype (measured: ValueError "kv_cache_dtype not
+# unify_kv_cache_spec_page_size — its separate fp32 scale tensor makes the
-# supported", crash loop). Pin the TURBOQUANT backend explicitly via the
+# per-token page bytes 4160 (SWA, head_dim 256) vs 8256 (full, head_dim
-# --attention-backend flag (this vLLM build removed the VLLM_ATTENTION_
+# 512), not an integer ratio. turboquant_k8v4 dies in backend validation —
-# BACKEND env var — it warns "Unknown vLLM environment variable" and
+# the TURBOQUANT backend rejects "partial multimodal token full attention"
-# ignores it). The TURBOQUANT backend accepts any head size and the
+# (the mm-prefix path Gemma4's vision/audio tokens require), and the
-# turboquant_* dtypes, and the Gemma4 MTP speculator carries the target's
+# VLLM_ATTENTION_BACKEND env var does not exist in this build (the pin
-# selected backend through to the draft layers, so the drafter follows.
+# goes through --attention-backend). fp8 is hardware-dead on sm_86
-ATTENTION_BACKEND=TURBOQUANT
+# (cvt.e4m3x2 is sm_89+). Revisit on a vLLM release that documents TurboQuant
+# mm-prefix support.
 MAX_NUM_SEQS=16
 MAX_NUM_BATCHED_TOKENS=16384
 SPEC_NUM_TOKENS=4

scripts/serve-frank-small.sh

@@ -44,7 +44,6 @@ fi
 : "${MAX_NUM_SEQS:=16}"
 : "${MAX_NUM_BATCHED_TOKENS:=16384}"
 : "${SPEC_NUM_TOKENS:=4}"
-: "${ATTENTION_BACKEND:=TURBOQUANT}"
 : "${ENFORCE_EAGER:=0}"
 
 # Fail loudly rather than half-loading a model into VRAM: a missing file here
@@ -67,7 +66,6 @@ VLLM_FLAGS=(
 --kv-cache-dtype "$KV_CACHE_DTYPE"
 --max-num-seqs "$MAX_NUM_SEQS"
 --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
---attention-backend "$ATTENTION_BACKEND"
 --speculative-config "$SPEC_CONFIG"
 --reasoning-parser gemma4
 --enable-auto-tool-choice

tests/frank_small_unit.bats

@@ -98,7 +98,7 @@ QUADLET_LARGE="$REPO_ROOT/dankclaw/frank-large/quadlet/frank-large.container"
     ! grep -q 'HALOGEN_VISION_TOWER' "$QUADLET_LARGE"
 }
 
-@test "frank-small serves the full window with turboquant KV on a dedicated card" {
+@test "frank-small serves the full window with bf16 KV on a dedicated card" {
     local maxlen kvdtype util seqs
     maxlen="$(sed -nE 's/^MAX_MODEL_LEN=([0-9]+)$/\1/p' "$ENV_FILE")"
     kvdtype="$(sed -nE 's/^KV_CACHE_DTYPE=(.+)$/\1/p' "$ENV_FILE")"
@@ -106,19 +106,17 @@ QUADLET_LARGE="$REPO_ROOT/dankclaw/frank-large/quadlet/frank-large.container"
     seqs="$(sed -nE 's/^MAX_NUM_SEQS=([0-9]+)$/\1/p' "$ENV_FILE")"
     # The model's FULL trained window (262144), not a divided one.
     [ "$maxlen" = "262144" ]
-    # TurboQuant k8v4 KV (K 8-bit, V 4-bit, in-tensor block scales,
+    # bf16 KV (auto): the only dtype that boots Gemma4 in this vLLM build.
-    # dedicated TURBOQUANT backend with an Ampere path). The int8/fp8
+    # int8_per_token_head dies in unify_kv_cache_spec_page_size (separate
-    # per-token-head dtypes cannot boot this model: their separate fp32
+    # fp32 scale tensor breaks the 256/512 head-dim page-size ratio),
-    # scale tensor breaks the page-size ratio between Gemma4's SWA
+    # turboquant_k8v4 dies in backend validation (no mm-prefix attention),
-    # (head_dim 256) and full-attention (head_dim 512) layers, and fp8 is
+    # and fp8 is hardware-dead on sm_86 — see frank-small.env.
-    # hardware-dead on sm_86 anyway.
+    [ "$kvdtype" = "auto" ]
-    [ "$kvdtype" = "turboquant_k8v4" ]
+    # No pinned attention backend: Gemma4's config hook selects TRITON_ATTN
-    # The TURBOQUANT backend must be pinned via --attention-backend:
+    # for the mixed head dims, and pinning any other backend breaks the KV
-    # Gemma4's config hook force-selects TRITON_ATTN (mixed 256/512 head
+    # dtype or mm-prefix validation.
-    # dims) whenever no backend is selected, and TRITON_ATTN rejects the
+    ! grep -q '^ATTENTION_BACKEND' "$ENV_FILE"
-    # turboquant dtype. This vLLM build has no attention-backend env var.
+    ! grep -q -- '--attention-backend' "$SERVE"
-    grep -q '^ATTENTION_BACKEND=TURBOQUANT$' "$ENV_FILE"
-    grep -q -- '--attention-backend "$ATTENTION_BACKEND"' "$SERVE"
     # The 3090 is dedicated to this service, so the engine may take most of
     # it; the KV pool, not the seq ceiling, caps full-context concurrency.
     [ "$util" = "0.95" ]