fix(dflash2): prefix-cache draft KV를 fail closed한다

launchers/start-glm53-nvfp4-tp4.sh

@@ -36,7 +36,7 @@ if [ -f "$PROFILE_ENV" ]; then
   for _v in IMAGE MOE_BACKEND ENABLE_EP EAGER GRAPH_CAP MAX_SEQS MAX_BATCHED MAX_LEN \
             GMU SPEC_K KV_DTYPE KV_BYTES DFLASH2 SPEC ASYNC_SCHED ATTN_BACKEND \
             MODEL_HOST_PATH SERVED_NAME DRAFT_TP DRAFT_KV CUSTOM_OPS_AXIS COMPILE_CFG \
-            EXTRA_ENV LOAD_FORMAT DRAFT_SAMPLE REJECT_METHOD $_vllm_keys; do
+            EXTRA_ENV LOAD_FORMAT DRAFT_SAMPLE REJECT_METHOD PREFIX_CACHE $_vllm_keys; do
     if [ -n "${!_v:-}" ]; then _caller="$_caller $_v=$(printf %q "${!_v}")"; fi
   done
   # shellcheck disable=SC1090
@@ -169,6 +169,21 @@ fi
 # num_speculative_tokens MUST be 7 (drafter block 8 minus the verified token).
 DFLASH2="${DFLASH2:-1}"
 DRAFT_HOST_PATH=/home/choiceoh/models/GLM-5.3-Flash-DFlash2
+# DFlash/DSpark synthesize their context KV from target hidden states. Tokens
+# restored by automatic prefix caching do not run through the target, so this
+# image leaves their draft KV slots unwritten while draft attention still
+# reads them. That is upstream vLLM #47926: long shared-prefix workloads can
+# collapse to position-0-only acceptance. Until that draft PR's four-file
+# state/block-table repair lands, fail closed and make every prompt token flow
+# through the target. PREFIX_CACHE=1 is an explicit throughput-first rollback;
+# it may restore TTFT reuse at the cost of DFlash2 acceptance on cache hits.
+PREFIX_CACHE="${PREFIX_CACHE:-0}"
+case "$PREFIX_CACHE" in
+  0) PREFIX_CACHE_FLAG="--no-enable-prefix-caching" ;;
+  1) PREFIX_CACHE_FLAG="--enable-prefix-caching" ;;
+  *) echo "ABORT: PREFIX_CACHE must be 0 or 1 (got $PREFIX_CACHE)" >&2; exit 2 ;;
+esac
+[ "$PREFIX_CACHE" = 1 ] || echo "prefix-cache: disabled for DFlash2 draft-KV safety"
 # The former AUDIT overlay replaced V1 files, but this image runs V2 Model
 # Runner.  Refuse the old switch instead of claiming an audit that cannot run.
 [ "${AUDIT:-0}" = 0 ] || {
@@ -465,6 +480,7 @@ SERVE_ARGS="$MODEL_PATH \
 ${ATTN_BACKEND:+--attention-backend $ATTN_BACKEND }\
 --max-model-len $MAX_LEN \
 --max-num-seqs $MAX_SEQS --max-num-batched-tokens $MAX_BATCHED --block-size 2304 --moe-backend $MOE_BACKEND \
+$PREFIX_CACHE_FLAG \
 --load-format $LOAD_FORMAT \
 ${EP_FLAG:+$EP_FLAG }\
 $SPECCFG_VAL \
@@ -483,7 +499,7 @@ $EAGER_FLAG --enable-flashinfer-autotune \
 if [ "${DRY_RUN:-0}" = 1 ]; then
   echo "profile   : ${PROFILE_ENV:-<none>}"
   for _k in IMAGE MOE_BACKEND ENABLE_EP VLLM_B12X_EP_COMPACT VLLM_B12X_EP_NO_DUMMY KV_DTYPE EAGER GRAPH_CAP GMU MAX_SEQS \
-            MAX_BATCHED MAX_LEN DFLASH2 SPEC SPEC_K ASYNC_SCHED \
+            MAX_BATCHED MAX_LEN DFLASH2 SPEC SPEC_K ASYNC_SCHED PREFIX_CACHE \
             DRAFT_SAMPLE REJECT_METHOD; do
     printf '  %-12s %s\n' "$_k" "${!_k:-<unset>}"
   done

profiles/glm53.env

@@ -166,6 +166,13 @@ VLLM_TARGET_LM_HEAD_FP8=1
 # Verdict: both deltas within +/-2 pct => precision retired; the gap then
 # lives in the drafter itself or bench content (dsv4's 78.5% came from its
 # own bench mix) -- DRAFT_TP and content-matched benches are the next axes.
+# DFlash2 context KV is built only for tokens that pass through the target.
+# vLLM enables automatic prefix caching by default, but cache-hit tokens skip
+# that forward and this image still lets the drafter attend to their unwritten
+# KV slots (upstream draft PR #47926). Prefer acceptance correctness over TTFT
+# reuse until that multi-file runtime fix lands. Set 1 only as an explicit
+# throughput-first rollback for a workload known not to reuse prefixes.
+PREFIX_CACHE=0
 # The drafter's block is 8; 7 is that minus the verified token. The launcher's
 # default of 4 contradicts its own comment.
 SPEC_K=7

tests/test_logic.py

@@ -5217,6 +5217,37 @@ def test_launcher_reject_method_gate() -> None:
     print("  launcher reject-method gate ... OK")
 
 
+def test_dflash2_prefix_cache_fail_closed() -> None:
+    """Cache-restored target tokens have no guaranteed DFlash context KV."""
+    launcher = open(
+        "launchers/start-glm53-nvfp4-tp4.sh", encoding="utf-8"
+    ).read()
+    profile = open("profiles/glm53.env", encoding="utf-8").read()
+
+    check("PREFIX_CACHE=0" in profile,
+          "the DFlash2 profile must prefer valid draft KV over prefix TTFT reuse")
+    check('PREFIX_CACHE="${PREFIX_CACHE:-0}"' in launcher,
+          "a profile-less DFlash2 launch must also fail closed")
+    check("ABORT: PREFIX_CACHE must be 0 or 1" in launcher,
+          "an invalid cache-safety value must not reach vLLM")
+    check('0) PREFIX_CACHE_FLAG="--no-enable-prefix-caching"' in launcher
+          and '1) PREFIX_CACHE_FLAG="--enable-prefix-caching"' in launcher,
+          "both explicit vLLM BooleanOptionalAction flags must be wired")
+    serve = launcher[launcher.index('SERVE_ARGS="'):]
+    check("$PREFIX_CACHE_FLAG" in serve,
+          "the validated prefix-cache decision must reach the serve command")
+    names = _launcher_caller_passthrough(launcher)
+    check("PREFIX_CACHE" in names,
+          "a caller must be able to make the documented throughput rollback")
+    dry_run = launcher[launcher.index('if [ "${DRY_RUN:-0}" = 1 ]'):]
+    check("PREFIX_CACHE" in dry_run,
+          "dry-run must expose whether draft-KV safety or prefix reuse won")
+    check(launcher.index('PREFIX_CACHE="${PREFIX_CACHE:-0}"')
+          < launcher.index('SERVE_ARGS="'),
+          "prefix-cache validation must happen before serve args are frozen")
+    print("  dflash2 prefix-cache fail-closed  OK")
+
+
 def test_accept_profile_conditional_arithmetic() -> None:
     """pos[i] is a MARGINAL count; the conditional is pos[i] / pos[i-1].
 
@@ -5282,6 +5313,7 @@ def test_accept_profile_conditional_arithmetic() -> None:
     test_hotpath_env_latches()
     test_launcher_load_format_gate()
     test_launcher_reject_method_gate()
+    test_dflash2_prefix_cache_fail_closed()
     test_accept_profile_conditional_arithmetic()
     test_launcher_nofile_limit()
     test_prefill_ladder_probe()