frank-small: --attention-backend flag (env var removed in this build)

The VLLM_ATTENTION_BACKEND env var does not exist in this vLLM dev
build (warns 'Unknown vLLM environment variable' and is ignored), so
the pin landed nowhere and Gemma4's hook force-selected TRITON_ATTN
again. Pass --attention-backend TURBOQUANT as a CLI flag instead.

scripts/frank-small.env

@@ -85,14 +85,16 @@ LISTEN_PORT=8000
 MAX_MODEL_LEN=262144
 GPU_MEM_UTIL=0.95
 KV_CACHE_DTYPE=turboquant_k8v4
-# Gemma4's config hook force-selects TRITON_ATTN for the mixed 256/512 head
+# ATTENTION_BACKEND: Gemma4's config hook force-selects TRITON_ATTN for the
-# dims, but only when no backend is selected — and TRITON_ATTN rejects the
+# mixed 256/512 head dims whenever no backend is selected, and TRITON_ATTN
-# turboquant KV dtype (measured: ValueError "kv_cache_dtype not supported",
+# rejects the turboquant KV dtype (measured: ValueError "kv_cache_dtype not
-# crash loop). Pin the TURBOQUANT backend explicitly: it accepts any head
+# supported", crash loop). Pin the TURBOQUANT backend explicitly via the
-# size, the turboquant dtypes, and has the Ampere path, and the Gemma4 MTP
+# --attention-backend flag (this vLLM build removed the VLLM_ATTENTION_
-# speculator carries the target's selected backend through to the draft
+# BACKEND env var — it warns "Unknown vLLM environment variable" and
-# layers, so the drafter follows.
+# ignores it). The TURBOQUANT backend accepts any head size and the
-VLLM_ATTENTION_BACKEND=TURBOQUANT
+# turboquant_* dtypes, and the Gemma4 MTP speculator carries the target's
+# selected backend through to the draft layers, so the drafter follows.
+ATTENTION_BACKEND=TURBOQUANT
 MAX_NUM_SEQS=16
 MAX_NUM_BATCHED_TOKENS=16384
 SPEC_NUM_TOKENS=4

scripts/serve-frank-small.sh

@@ -44,6 +44,7 @@ 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
@@ -66,6 +67,7 @@ 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

@@ -113,10 +113,12 @@ QUADLET_LARGE="$REPO_ROOT/dankclaw/frank-large/quadlet/frank-large.container"
     # (head_dim 256) and full-attention (head_dim 512) layers, and fp8 is
     # hardware-dead on sm_86 anyway.
     [ "$kvdtype" = "turboquant_k8v4" ]
-    # The TURBOQUANT backend must be pinned explicitly: Gemma4's config
+    # The TURBOQUANT backend must be pinned via --attention-backend:
-    # hook force-selects TRITON_ATTN (mixed 256/512 head dims) whenever no
+    # Gemma4's config hook force-selects TRITON_ATTN (mixed 256/512 head
-    # backend is selected, and TRITON_ATTN rejects the turboquant dtype.
+    # dims) whenever no backend is selected, and TRITON_ATTN rejects the
-    grep -q '^VLLM_ATTENTION_BACKEND=TURBOQUANT$' "$ENV_FILE"
+    # turboquant dtype. This vLLM build has no attention-backend env var.
+    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" ]