From fa63bb9db6f48108077fb5497081f7189092d163 Mon Sep 17 00:00:00 2001 From: Mike G <180722391+mikekg@users.noreply.github.com> Date: Mon, 15 Jun 2026 06:49:57 -0700 Subject: [PATCH] Remove redundant Triton KV cache dtype asserts and enforce architectural support (fp8 >= sm89) (#43914) Signed-off-by: Mike G <180722391+mikekg@users.noreply.github.com> Co-authored-by: Michael Gschwind --- vllm/v1/attention/backends/triton_attn.py | 20 +++++++ .../ops/triton_reshape_and_cache_flash.py | 52 +++++-------------- 2 files changed, 33 insertions(+), 39 deletions(-) diff --git a/vllm/v1/attention/backends/triton_attn.py b/vllm/v1/attention/backends/triton_attn.py index 377e9e7ab1..6c67735e9f 100644 --- a/vllm/v1/attention/backends/triton_attn.py +++ b/vllm/v1/attention/backends/triton_attn.py @@ -464,6 +464,26 @@ class TritonAttentionImpl(AttentionImpl): else: self.sliding_window = (sliding_window - 1, 0) self.kv_cache_dtype = kv_cache_dtype + cap = current_platform.get_device_capability() + cap_str = cap.as_version_str() if cap is not None else "unknown" + dev = current_platform.get_device_name() + if self.kv_cache_dtype.startswith("fp8") and not ( + current_platform.has_device_capability(89) + ): + suggested = "float16" if (cap is None or cap.to_int() < 80) else "bfloat16" + raise ValueError( + f"FP8 KV cache is not supported by the Triton attention backend " + f"on {dev} (compute capability {cap_str}); native FP8 (fp8e4nv) " + f"requires SM89+. Re-run with --kv-cache-dtype {suggested}." + ) + if self.kv_cache_dtype == "bfloat16" and not ( + current_platform.has_device_capability(80) + ): + raise ValueError( + f"bfloat16 KV cache is not supported on {dev} (compute capability " + f"{cap_str}); bfloat16 requires SM80+. Re-run with " + f"--kv-cache-dtype float16." + ) if logits_soft_cap is None: # In flash-attn, setting logits_soft_cap as 0 means no soft cap. logits_soft_cap = 0 diff --git a/vllm/v1/attention/ops/triton_reshape_and_cache_flash.py b/vllm/v1/attention/ops/triton_reshape_and_cache_flash.py index 08c6673fb5..3959cba575 100644 --- a/vllm/v1/attention/ops/triton_reshape_and_cache_flash.py +++ b/vllm/v1/attention/ops/triton_reshape_and_cache_flash.py @@ -17,9 +17,16 @@ _NATIVE_KV_CACHE_DTYPES = {"auto", "float16", "bfloat16", "float32", "half", "fl def _is_supported_kv_cache_dtype(kv_cache_dtype: str) -> bool: - return kv_cache_dtype in _NATIVE_KV_CACHE_DTYPES or is_quantized_kv_cache( - kv_cache_dtype - ) + if not ( + kv_cache_dtype in _NATIVE_KV_CACHE_DTYPES + or is_quantized_kv_cache(kv_cache_dtype) + ): + return False + if kv_cache_dtype.startswith("fp8"): + return current_platform.has_device_capability(89) + if kv_cache_dtype == "bfloat16": + return current_platform.has_device_capability(80) + return True @triton.jit @@ -359,7 +366,9 @@ def triton_reshape_and_cache_flash( page_stride = key_cache.stride()[1] assert _is_supported_kv_cache_dtype(kv_cache_dtype), ( - f"unsupported kv_cache_dtype (str), got {kv_cache_dtype}." + f"Triton reshape-and-cache cannot store kv_cache_dtype={kv_cache_dtype} " + f"on this device: an FP8 KV cache needs native fp8e4nv (SM89+). Use " + f"--kv-cache-dtype bfloat16 (or float16 on SM75)." ) kv_cache_torch_dtype = ( current_platform.fp8_dtype() @@ -374,23 +383,7 @@ def triton_reshape_and_cache_flash( # (e.g. explicit cast to fp8e4m3fnuz is not supported in triton 3.4) key_cache = key_cache.view(kv_cache_torch_dtype) value_cache = value_cache.view(kv_cache_torch_dtype) - assert kv_cache_dtype != torch.uint8, ( - "explicit fp8 cast and store to " - "uint8 is not supported by triton reshape_and_cache_flash" - ) - FP8_KV_CACHE = is_quantized_kv_cache(kv_cache_dtype) - assert (not FP8_KV_CACHE) or kv_cache_torch_dtype in [ - torch.float8_e4m3fn, - torch.float8_e5m2, - torch.uint8, - torch.float8_e4m3fnuz, - ], ( - "unsupported dtype of KV cache tensor, got " - "{kv_cache_torch_dtype}. Supported kv cache dtypes: fp8e4m3fn, " - "fp8e5m2, uint8, bfloat16, float16, float32, fp8e4m3fnuz." - ) - # heuristics instead of autotuning TILE_SIZE = min(2048, triton.next_power_of_2(n)) if current_platform.is_rocm() or current_platform.is_xpu(): @@ -537,9 +530,6 @@ def triton_reshape_and_cache_flash_diffkv( block_stride = kv_cache.stride()[0] page_stride = kv_cache.stride()[1] - assert _is_supported_kv_cache_dtype(kv_cache_dtype), ( - f"unsupported kv_cache_dtype (str), got {kv_cache_dtype}." - ) kv_cache_torch_dtype = ( current_platform.fp8_dtype() if is_quantized_kv_cache(kv_cache_dtype) @@ -550,23 +540,7 @@ def triton_reshape_and_cache_flash_diffkv( # to avoid erounous implicit cast in triton kernel (tl.store to uint8) # (e.g. explicit cast to fp8e4m3fnuz is not supported in triton 3.4) kv_cache = kv_cache.view(kv_cache_torch_dtype) - assert kv_cache_dtype != torch.uint8, ( - "explicit fp8 cast and store to " - "uint8 is not supported by triton reshape_and_cache_flash_diffkv" - ) - FP8_KV_CACHE = is_quantized_kv_cache(kv_cache_dtype) - assert (not FP8_KV_CACHE) or kv_cache_torch_dtype in [ - torch.float8_e4m3fn, - torch.float8_e5m2, - torch.uint8, - torch.float8_e4m3fnuz, - ], ( - "unsupported dtype of KV cache tensor, got " - "{kv_cache_torch_dtype}. Supported kv cache dtypes: fp8e4m3fn, " - "fp8e5m2, uint8, bfloat16, float16, float32, fp8e4m3fnuz." - ) - # heuristics instead of autotuning TILE_SIZE = max(head_size_k, head_size_v) TILE_SIZE = triton.next_power_of_2(TILE_SIZE)