[Docs] Point users at --disable-flashinfer-autotune (#61489)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

benchmarks/replayssm/e2e_decode_speedup.py

@@ -120,7 +120,7 @@ def run_worker(args):
     if args.disable_flashinfer_autotune:
         # FP4-MoE autotuner is unstable under CUDA-graph capture on Blackwell;
         # re-enable (--no-disable-flashinfer-autotune) only for non-FP4 models.
-        llm_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+        llm_kwargs["enable_flashinfer_autotune"] = False
     if mode == "replayssm":
         llm_kwargs.update(use_replayssm=True, replayssm_buffer_len=args.buffer_len)
 

docs/design/optimization_levels.md

@@ -37,7 +37,7 @@ Settings:
 - `-cc.cudagraph_mode=NONE`
 - `-cc.mode=NONE` (also resulting in `-cc.custom_ops=["none"]`)
 - `-cc.pass_config.fuse_...=False` (all fusions disabled)
-- `--kernel-config.enable_flashinfer_autotune=False`
+- `--disable-flashinfer-autotune`
 
 ### `-O1`: Fast Optimization
 
@@ -49,7 +49,7 @@ Settings:
 
 - `-cc.cudagraph_mode=PIECEWISE`
 - `-cc.mode=VLLM_COMPILE`
-- `--kernel-config.enable_flashinfer_autotune=True`
+- `--enable-flashinfer-autotune`
 
 Fusions:
 
@@ -84,6 +84,6 @@ in the future that are more time-consuming or experimental.
 
 ### Common Issues
 
-1. **Startup Time Too Long**: Use `-O0` or `-O1` for faster startup
+1. **Startup Time Too Long**: Use `-O0` or `-O1` for faster startup, or pass `--disable-flashinfer-autotune` to skip only FlashInfer autotuning
 2. **Compilation Errors**: Use `debug_dump_path` for additional debugging information
 3. **Performance Issues**: Ensure using `-O2` for production

recipes/models/MiniMaxAI/MiniMax-M3.yaml

@@ -247,7 +247,7 @@ variants:
           - "16384"
           - "--gpu-memory-utilization"
           - "0.9"
-          - "--no-enable-flashinfer-autotune"
+          - "--disable-flashinfer-autotune"
           - "--stream-interval"
           - "20"
         extra_env:
@@ -829,7 +829,7 @@ guide: |
     --speculative-config '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASHINFER","use_local_argmax_reduction":true}' \
     --max-model-len 1048576 --max-num-batched-tokens 16384 --max-num-seqs <S> \
     --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[<sizes from the table>,256,512,1024,2048]}' \
-    --gpu-memory-utilization 0.9 --no-enable-flashinfer-autotune --stream-interval 20 --language-model-only \
+    --gpu-memory-utilization 0.9 --disable-flashinfer-autotune --stream-interval 20 --language-model-only \
     --kv-events-config '{"publisher":"zmq","topic":"kv-events","enable_kv_cache_events":true,"endpoint":"tcp://*:5200"}' \
     --reasoning-parser minimax_m3 --dyn-reasoning-parser minimax_m3 --dyn-tool-call-parser minimax_m3
   ```

recipes/models/deepseek-ai/DeepSeek-V3.2.yaml

@@ -24,7 +24,7 @@ model:
   supports_dcp: true
   base_args:
     - "--trust-remote-code"
-    - "--kernel-config.enable_flashinfer_autotune=False"
+    - "--disable-flashinfer-autotune"
   base_env: {}
 
 dependencies:
@@ -122,7 +122,7 @@ guide: |
   vllm serve deepseek-ai/DeepSeek-V3.2 \
     --tensor-parallel-size 8 \
     --trust-remote-code \
-    --kernel-config.enable_flashinfer_autotune=False \
+    --disable-flashinfer-autotune \
     --tokenizer-mode deepseek_v32 \
     --tool-call-parser deepseek_v32 \
     --enable-auto-tool-choice \

recipes/models/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp.yaml

@@ -271,7 +271,7 @@ guide: |
   - `VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS` excludes the FP4 block-scale MoE op from
     FlashInfer's startup autotuning. Export it if autotuning that op fails or costs more
     startup time than it earns back on your FlashInfer build; it is a tuning knob, not a
-    requirement. The Advanced row's `--no-enable-flashinfer-autotune` turns autotuning
+    requirement. The Advanced row's `--disable-flashinfer-autotune` turns autotuning
     off wholesale instead.
   - `--max-model-len 32768` bounds KV cache for the benchmark. The checkpoint advertises
     1M tokens; that is **not** what was measured, and 1M-token KV needs far more memory

recipes/models/deepseek-ai/DeepSeek-V4-Flash.yaml

@@ -591,7 +591,7 @@ strategy_overrides:
     hardware_overrides:
       hopper:
         extra_args:
-          - "--no-enable-flashinfer-autotune"
+          - "--disable-flashinfer-autotune"
       blackwell:
     # Replaces recipe-level blackwell override: keeps the FP4 indexer cache
     # but drops --moe-backend deep_gemm_mega_moe (TP-only path uses the
@@ -599,7 +599,7 @@ strategy_overrides:
         extra_args:
          - "--attention_config.indexer_kv_dtype"
          - "mxfp4"
-         - "--no-enable-flashinfer-autotune"
+         - "--disable-flashinfer-autotune"
       rtx_pro_6000_8x:
         extra_args:
           - "--attention_config.indexer_kv_dtype"

recipes/models/deepseek-ai/DeepSeek-V4-Pro.yaml

@@ -183,7 +183,7 @@ hardware_overrides:
       - "0.95"
       - "--max-num-seqs"
       - "512"
-      - "--no-enable-flashinfer-autotune"
+      - "--disable-flashinfer-autotune"
       - "--compilation-config"
       - '{"mode": 0, "cudagraph_mode": "FULL_DECODE_ONLY"}'
   blackwell:
@@ -214,14 +214,14 @@ strategy_overrides:
     hardware_overrides:
       hopper:
         extra_args:
-          - "--no-enable-flashinfer-autotune"
+          - "--disable-flashinfer-autotune"
       blackwell:
         # Replaces recipe-level blackwell override: keeps the FP4 indexer cache
         # but drops any variant-level --moe-backend selection (TP-only uses
         # the default MoE backend).
         extra_args:
           - "--attention_config.indexer_kv_dtype=mxfp4"
-          - "--no-enable-flashinfer-autotune"
+          - "--disable-flashinfer-autotune"
   single_node_tep:
     extra_args:
       - "--compilation-config"
@@ -485,7 +485,7 @@ guide: |
   ```text
   --tokenizer-mode deepseek_v4 --trust-remote-code
   --kv-cache-dtype fp8 --block-size 256
-  --no-disable-hybrid-kv-cache-manager --no-enable-flashinfer-autotune
+  --no-disable-hybrid-kv-cache-manager --disable-flashinfer-autotune
   --attention-config '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
   --speculative-config '{"method":"dspark","num_speculative_tokens":6,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
   ```
@@ -675,7 +675,7 @@ guide: |
     --numa-bind --numa-bind-nodes 0 0 1 1
     --tokenizer-mode deepseek_v4 --trust-remote-code
     --max-model-len 1048576 --kv-cache-dtype fp8 --block-size 256
-    --no-disable-hybrid-kv-cache-manager --no-enable-flashinfer-autotune
+    --no-disable-hybrid-kv-cache-manager --disable-flashinfer-autotune
     --attention-config '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
     --speculative-config '{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
   )

recipes/models/moonshotai/Kimi-K3.yaml

@@ -289,7 +289,7 @@ hardware_overrides:
       - "--moe-backend"
       - "marlin"
       - "--disable-custom-all-reduce"
-      - "--no-enable-flashinfer-autotune"
+      - "--disable-flashinfer-autotune"
       - "--max-num-batched-tokens"
       - "4096"
       - "--attention-backend"
@@ -303,7 +303,7 @@ hardware_overrides:
     extra_args:
       - "--load-format"
       - "fastsafetensors"
-      - "--no-enable-flashinfer-autotune"
+      - "--disable-flashinfer-autotune"
       - "--max-model-len"
       - "1048576"
       - "--kv-cache-dtype"
@@ -638,7 +638,7 @@ strategy_overrides:
         - "--max-num-batched-tokens"
         - "16384"
         - "--no-disable-hybrid-kv-cache-manager"
-        - "--no-enable-flashinfer-autotune"
+        - "--disable-flashinfer-autotune"
       env:
         VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT: "1"
     decode:
@@ -841,7 +841,7 @@ guide: |
     --compilation-config "{\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"cudagraph_capture_sizes\":[$CAPTURES]}" \
     --speculative-config "{\"model\":\"Inferact/Kimi-K3-DSpark\",\"attention_backend\":\"TOKENSPEED_MLA\",\"method\":\"dspark\",\"num_speculative_tokens\":$K,\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}" \
     --enable-prefix-caching --prefix-match-unit 128 \
-    --no-enable-flashinfer-autotune \
+    --disable-flashinfer-autotune \
     --moe-backend auto --load-format fastsafetensors \
     --stream-interval 10 --language-model-only --trust-remote-code \
     --enable-auto-tool-choice --tool-call-parser kimi_k3 --reasoning-parser kimi_k3 \

recipes/models/thinkingmachines/Inkling-Small.yaml

@@ -79,7 +79,7 @@ variants:
       h200: 2
     description: "NVFP4 checkpoint requiring at least 180 GB aggregated VRAM: W4A4 with TP1 on B300/GB300 or TP2 on B200/GB200, and W4A16 with TP4 on H100 or TP2 on H200."
     extra_args:
-      - "--kernel-config.enable_flashinfer_autotune=False"
+      - "--disable-flashinfer-autotune"
     extra_env:
       FLASH_ATTENTION_CUTE_DSL_CACHE_ENABLED: "1"
   mxfp4:
@@ -163,7 +163,7 @@ variants:
     hardware_overrides:
       nvidia:
         extra_args:
-          - "--kernel-config.enable_flashinfer_autotune=False"
+          - "--disable-flashinfer-autotune"
         extra_env:
           FLASH_ATTENTION_CUTE_DSL_CACHE_ENABLED: "1"
       mi300x:
@@ -309,7 +309,7 @@ guide: |
     --tool-call-parser inkling \
     --enable-auto-tool-choice \
     --tensor-parallel-size 1 \
-    --kernel-config.enable_flashinfer_autotune=False \
+    --disable-flashinfer-autotune \
     --trust-remote-code
   ```
 

recipes/models/thinkingmachines/Inkling.yaml

@@ -75,7 +75,7 @@ variants:
     vram_minimum_gb: 651
     description: "Weight-only NVFP4 (group size 16) on the routed MoE expert matmuls only; attention, norms, sconv, shared experts, routers, embeddings, and the layer 0-1 dense MLPs stay BF16. ~542 GB on disk. On Hopper, the MoE kernel dequantizes the NVFP4 expert weights to BF16 on the fly."
     extra_args:
-      - "--kernel-config.enable_flashinfer_autotune=False"
+      - "--disable-flashinfer-autotune"
     extra_env:
       FLASH_ATTENTION_CUTE_DSL_CACHE_ENABLED: "1"
   mxfp4:
@@ -143,7 +143,7 @@ variants:
     vram_minimum_gb: 2400
     description: "Full BF16 checkpoint (~1.8 TB on disk). Runs on Hopper and Blackwell."
     extra_args:
-      - "--kernel-config.enable_flashinfer_autotune=False"
+      - "--disable-flashinfer-autotune"
     extra_env:
       FLASH_ATTENTION_CUTE_DSL_CACHE_ENABLED: "1"
 
@@ -292,7 +292,7 @@ guide: |
     --tool-call-parser inkling \
     --enable-auto-tool-choice \
     --tensor-parallel-size 4 \
-    --kernel-config.enable_flashinfer_autotune=False \
+    --disable-flashinfer-autotune \
     --trust-remote-code
   ```
 

recipes/models/zai-org/GLM-5.3-Flash.yaml

@@ -135,7 +135,7 @@ strategy_hardware:
 hardware_overrides:
   hopper:
     extra_args:
-      - "--no-enable-flashinfer-autotune"
+      - "--disable-flashinfer-autotune"
   blackwell:
     # FP8 KV cache only works on Blackwell for this model; Hopper runs BF16 KV.
     extra_args:

recipes/models/zai-org/GLM-5.3.yaml

@@ -150,7 +150,7 @@ variants:
           - "w2_weight"
           - "--load-format"
           - "fastsafetensors"
-          - "--no-enable-flashinfer-autotune"
+          - "--disable-flashinfer-autotune"
           - "--moe-backend"
           - "flashinfer_trtllm"
   w4a8:

tests/compile/fusions_e2e/test_tp1_quant.py

@@ -103,7 +103,7 @@ def test_tp1_fp8_fusions(
     model_kwargs["hf_overrides"] = hf_overrides(n_layers)
     model_kwargs["load_format"] = "dummy"
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
 
     compilation_config = dict(
         use_inductor_graph_partition=inductor_graph_partition,
@@ -180,7 +180,7 @@ def test_tp1_fp4_fusions(
     model_kwargs["hf_overrides"] = hf_overrides(n_layers)
     model_kwargs["load_format"] = "dummy"
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
 
     compilation_config = dict(
         use_inductor_graph_partition=inductor_graph_partition,
@@ -226,7 +226,7 @@ def test_tp1_fp4_mixed_precision_act_fusion(
     model_kwargs["hf_overrides"] = hf_overrides(n_layers)
     model_kwargs["load_format"] = "dummy"
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
 
     compilation_config = dict(
         use_inductor_graph_partition=inductor_graph_partition,

tests/compile/fusions_e2e/test_tp2_ar_rms.py

@@ -109,7 +109,7 @@ def test_tp2_ar_rms_fp8_fusions(
     model_kwargs["hf_overrides"] = hf_overrides(n_layers)
     model_kwargs["load_format"] = "dummy"
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
     model_kwargs["disable_custom_all_reduce"] = False
     if current_platform.is_rocm() and (
         "llama-3" in model_name_lower or "qwen" in model_name_lower
@@ -202,7 +202,7 @@ def test_tp2_ar_rms_fp4_fusions(
     model_kwargs["hf_overrides"] = hf_overrides(n_layers)
     model_kwargs["load_format"] = "dummy"
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
     model_kwargs["disable_custom_all_reduce"] = False
 
     compilation_config = dict(
@@ -280,7 +280,7 @@ def test_tp2_ar_rms_fusions(
     model_kwargs["load_format"] = "dummy"
     model_kwargs["model_impl"] = model_impl
     model_kwargs["max_model_len"] = 1024
-    model_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    model_kwargs["enable_flashinfer_autotune"] = False
     model_kwargs["disable_custom_all_reduce"] = False
 
     compilation_config = dict(

tests/evals/gsm8k/configs/Kimi-K3-pruned75-DSpark-TP4.yaml

@@ -9,7 +9,7 @@ server_args: >-
   --gpu-memory-utilization 0.90
   --tensor-parallel-size 4
   --load-format fastsafetensors
-  --no-enable-flashinfer-autotune
+  --disable-flashinfer-autotune
   --max-model-len 1048576
   --max-num-batched-tokens 16384
   --max-num-seqs 128

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-int6wc-hadamard-humming-act-int8.yaml

@@ -8,7 +8,7 @@ server_args: >-
   --max-model-len 8192
   --tensor-parallel-size 1
   --quantization humming
-  --kernel-config.enable_flashinfer_autotune=False
+  --disable-flashinfer-autotune
 env:
   VLLM_HUMMING_ONLINE_QUANT_CONFIG: '{"dtype":"int6","hadamard_block_size":-1}'
   VLLM_HUMMING_INPUT_QUANT_CONFIG: '{"dtype":"int8"}'

tests/evals/gsm8k/configs/humming/Qwen3.5-35B-A3B-int6wc-hadamard-humming.yaml

@@ -8,6 +8,6 @@ server_args: >-
   --max-model-len 8192
   --tensor-parallel-size 1
   --quantization humming
-  --kernel-config.enable_flashinfer_autotune=False
+  --disable-flashinfer-autotune
 env:
   VLLM_HUMMING_ONLINE_QUANT_CONFIG: '{"dtype":"int6","hadamard_block_size":-1}'

tests/evals/gsm8k/configs/humming/Qwen3.5-9B-int6wc-hadamard-humming-act-int8.yaml

@@ -7,7 +7,7 @@ server_args: >-
   --enforce-eager
   --max-model-len 8192
   --quantization humming
-  --kernel-config.enable_flashinfer_autotune=False
+  --disable-flashinfer-autotune
 env:
   VLLM_HUMMING_ONLINE_QUANT_CONFIG: '{"dtype":"int6","hadamard_block_size":-1}'
   VLLM_HUMMING_INPUT_QUANT_CONFIG: '{"dtype":"int8"}'

tests/evals/gsm8k/configs/humming/Qwen3.5-9B-int6wc-hadamard-humming.yaml

@@ -7,6 +7,6 @@ server_args: >-
   --enforce-eager
   --max-model-len 8192
   --quantization humming
-  --kernel-config.enable_flashinfer_autotune=False
+  --disable-flashinfer-autotune
 env:
   VLLM_HUMMING_ONLINE_QUANT_CONFIG: '{"dtype":"int6","hadamard_block_size":-1}'

tests/models/kimi_k3/test_prefix_cache.py

@@ -44,7 +44,7 @@ BASE_ARGS = [
     "--language-model-only",
     "--load-format",
     "dummy",
-    "--no-enable-flashinfer-autotune",
+    "--disable-flashinfer-autotune",
     "--enable-prompt-tokens-details",
     # Default max_num_seqs OOMs one GPU during warmup.
     "--max-num-seqs",

tests/v1/e2e/general/test_sharded_sampling.py

@@ -113,7 +113,7 @@ def _generate(
         gpu_memory_utilization=0.7,
         # Timing-based kernel selection is the dominant source of cross-boot
         # numeric noise; disable it so the comparison bounds stay tight.
-        kernel_config={"enable_flashinfer_autotune": False},
+        enable_flashinfer_autotune=False,
     )
     try:
         # Guard against a vacuous comparison: verify every worker is in the

tests/v1/e2e/spec_decode/test_sharded_sampling.py

@@ -61,7 +61,7 @@ def test_mtp_sharded_sampling_equivalence(monkeypatch: pytest.MonkeyPatch):
                 max_num_seqs=len(test_prompts),
                 # Timing-based kernel selection is the dominant source of
                 # cross-boot numeric noise; disable for a tight comparison.
-                kernel_config={"enable_flashinfer_autotune": False},
+                enable_flashinfer_autotune=False,
             )
             try:
                 # Guard against a vacuous comparison: verify every worker is

vllm/config/kernel.py

@@ -278,7 +278,8 @@ class KernelConfig:
     """
 
     enable_flashinfer_autotune: bool = None  # type: ignore[assignment]
-    """If True, run FlashInfer autotuning during kernel warmup."""
+    """Benchmark FlashInfer kernel implementations during warmup and use the
+    fastest. Enabled by default except at `-O0`; disable for faster startup."""
 
     enable_rocm_segmented_attn_autotune: bool = False
     """If True, autotune ROCm segmented attention during kernel warmup on RDNA GPUs."""

vllm/model_executor/warmup/kernel_warmup.py

@@ -461,7 +461,8 @@ def flashinfer_autotune(runner: "GPUModelRunner", *, skip_attn: bool = False) ->
         cache_root = resolve_flashinfer_autotune_v2_root()
         if is_leader:
             logger.info_once(
-                "Using FlashInfer managed autotune cache (root=%s)",
+                "Using FlashInfer managed autotune cache (root=%s). Pass "
+                "--disable-flashinfer-autotune to skip autotuning.",
                 cache_root if cache_root is not None else "flashinfer default",
             )
         autotune_context = autotune_v2(
@@ -475,7 +476,11 @@ def flashinfer_autotune(runner: "GPUModelRunner", *, skip_attn: bool = False) ->
             f"{cache_path.stem}_dp{dp_rank}_rank{world.rank_in_group}{cache_path.suffix}"
         )
         if is_leader:
-            logger.info_once("Using FlashInfer autotune cache file: %s", cache_path)
+            logger.info_once(
+                "Using FlashInfer autotune cache file: %s. Pass "
+                "--disable-flashinfer-autotune to skip autotuning.",
+                cache_path,
+            )
 
         if _all_ranks_have_matching_cache(cache_path, tune_group):
             loaded = tuner.load_configs(str(cache_path))