[CPU] Support FP8 KV cache (#14741)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

csrc/cpu/cache.cpp

@@ -3,6 +3,12 @@
 
 #include "cpu_types.hpp"
 
+#if defined(__x86_64__)
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2
+#else
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES
+#endif
+
 namespace {
 template <typename scalar_t>
 void copy_blocks_cpu_impl(std::vector<torch::Tensor> const& key_caches,
@@ -95,8 +101,7 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
   }
 
   const int element_num_per_block = key_caches[0][0].numel();
-  VLLM_DISPATCH_FLOATING_TYPES(
+  DISPATCH_MACRO(key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
-      key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
     CPU_KERNEL_GUARD_IN(copy_blocks_cpu_impl)
     copy_blocks_cpu_impl<scalar_t>(key_caches, value_caches, block_mapping,
                                    element_num_per_block, num_layers);
@@ -118,14 +123,13 @@ void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
   int key_stride = key.stride(0);
   int value_stride = value.stride(0);
 
-  VLLM_DISPATCH_FLOATING_TYPES(
+  DISPATCH_MACRO(key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
-      key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
     CPU_KERNEL_GUARD_IN(reshape_and_cache_cpu_impl)
     reshape_and_cache_cpu_impl<scalar_t>(
         key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
         key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
-            slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride,
+        slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride, value_stride,
-            value_stride, num_heads, head_size, block_size, x);
+        num_heads, head_size, block_size, x);
     CPU_KERNEL_GUARD_OUT(reshape_and_cache_cpu_impl)
   });
 }

csrc/cpu/cpu_types_x86.hpp

@@ -16,9 +16,18 @@ namespace vec_op {
   AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
 
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES_FP8(...)        \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Float8_e5m2, __VA_ARGS__)
+
 #define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
+#define VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME,                                \
+                     VLLM_DISPATCH_CASE_FLOATING_TYPES_FP8(__VA_ARGS__))
+
 #ifndef CPU_OP_GUARD
   #define CPU_KERNEL_GUARD_IN(NAME)
   #define CPU_KERNEL_GUARD_OUT(NAME)

docs/source/getting_started/installation/cpu.md

@@ -189,7 +189,7 @@ vLLM CPU backend supports the following vLLM features:
 - Model Quantization (`INT8 W8A8, AWQ, GPTQ`)
 - Chunked-prefill
 - Prefix-caching
-- FP8-E5M2 KV-Caching (TODO)
+- FP8-E5M2 KV cache
 
 ## Related runtime environment variables
 

tests/basic_correctness/test_chunked_prefill.py

@@ -266,7 +266,7 @@ def test_with_prefix_caching(
 
 
 @pytest.mark.parametrize("model", ["facebook/opt-125m"])
-@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("dtype", ["bfloat16", "half"])
 @pytest.mark.parametrize("max_tokens", [32])
 @pytest.mark.parametrize("chunked_prefill_token_size", [1, 4, 16])
 @pytest.mark.parametrize("enforce_eager", [False])
@@ -303,7 +303,7 @@ def test_models_cpu(
 @pytest.mark.parametrize("max_tokens", [16])
 @pytest.mark.parametrize("enforce_eager", [False])
 @pytest.mark.parametrize("chunk_size", [30, 32])
-@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("dtype", ["bfloat16", "half"])
 @pytest.mark.cpu_model
 @pytest.mark.skipif(not current_platform.is_cpu(), reason="CPU only")
 def test_with_prefix_caching_cpu(

tests/models/decoder_only/language/test_fp8.py

@@ -11,6 +11,7 @@ import pytest
 
 from tests.kernels.utils import override_backend_env_variable
 from tests.quantization.utils import is_quant_method_supported
+from vllm.platforms import current_platform
 
 from ...utils import check_logprobs_close
 
@@ -93,3 +94,63 @@ def test_models(
         name_0="fp16_kv_cache",
         name_1="fp8_kv_cache",
     )
+
+
+@pytest.mark.cpu_model
+@pytest.mark.skipif(not current_platform.is_cpu(),
+                    reason="test for the CPU backend.")
+@pytest.mark.parametrize(
+    "kv_cache_dtype,base_model,test_model",
+    [
+        # Test BF16 checkpoint w. fp8_e5m2 kv-cache.
+        ("fp8_e5m2", "meta-llama/Llama-3.2-1B-Instruct",
+         "meta-llama/Llama-3.2-1B-Instruct"),
+    ])
+# Due to low-precision numerical divergence, we only test logprob of 4 tokens
+@pytest.mark.parametrize("max_tokens", [4])
+# Due to low-precision numerical divergence, this test is too sensitive for
+# the async postprocessor
+@pytest.mark.parametrize("disable_async_output_proc", [True])
+def test_cpu_models(
+    vllm_runner,
+    example_prompts,
+    kv_cache_dtype: str,
+    base_model: str,
+    test_model: str,
+    max_tokens: int,
+    disable_async_output_proc: bool,
+) -> None:
+    """
+    Only checks log probs match to cover the discrepancy in
+    numerical sensitive kernels.
+    """
+
+    MAX_MODEL_LEN = 1024
+    NUM_LOG_PROBS = 8
+
+    with vllm_runner(
+            base_model,
+            max_model_len=MAX_MODEL_LEN,
+            dtype="bfloat16",
+            kv_cache_dtype="auto",
+            disable_async_output_proc=disable_async_output_proc,
+    ) as vllm_model:
+        baseline_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, NUM_LOG_PROBS)
+
+    with vllm_runner(
+            test_model,
+            max_model_len=MAX_MODEL_LEN,
+            dtype="bfloat16",
+            kv_cache_dtype=kv_cache_dtype,
+            disable_async_output_proc=disable_async_output_proc,
+    ) as vllm_model:
+        test_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts, max_tokens, NUM_LOG_PROBS)
+
+    check_logprobs_close(
+        outputs_0_lst=baseline_outputs,
+        outputs_1_lst=test_outputs,
+        name_0="bf16_kv_cache",
+        name_1="fp8_kv_cache",
+    )

vllm/attention/backends/torch_sdpa.py

@@ -17,7 +17,7 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               is_quantized_kv_cache)
 # yapf: enable
 from vllm.attention.backends.utils import CommonAttentionState
-from vllm.attention.ops.ipex_attn import PagedAttention
+from vllm.attention.ops.ipex_attn import PagedAttention, _use_ipex
 from vllm.attention.ops.paged_attn import PagedAttentionMetadata
 from vllm.logger import init_logger
 from vllm.utils import make_tensor_with_pad
@@ -431,10 +431,11 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
             raise ValueError(
                 f"Head size {head_size} is not supported by PagedAttention. "
                 f"Supported head sizes are: {supported_head_sizes}.")
-        if is_quantized_kv_cache(kv_cache_dtype):
+
+        if is_quantized_kv_cache(kv_cache_dtype) and not _use_ipex:
             raise NotImplementedError(
-                "Torch SDPA backend does not support FP8 KV cache. "
+                "Torch SDPA backend FP8 KV cache requires "
-                "Please use xFormers backend instead.")
+                "intel_extension_for_pytorch support.")
         self.attn_type = attn_type
 
     def forward(

vllm/platforms/cpu.py

@@ -60,9 +60,6 @@ class CpuPlatform(Platform):
         # Reminder: Please update docs/source/features/compatibility_matrix.md
         # If the feature combo become valid
         if not model_config.enforce_eager:
-            logger.warning(
-                "CUDA graph is not supported on CPU, fallback to the eager "
-                "mode.")
             model_config.enforce_eager = True
 
         cache_config = vllm_config.cache_config
@@ -70,6 +67,25 @@ class CpuPlatform(Platform):
         if cache_config and cache_config.block_size is None:
             cache_config.block_size = 16
 
+        scheduler_config = vllm_config.scheduler_config
+        if ((scheduler_config.chunked_prefill_enabled
+             or cache_config.enable_prefix_caching)
+                and cache_config.cache_dtype != "auto"):
+            raise RuntimeError("Chunked-prefill and prefix-cache on the CPU "
+                               "backend is not compatible with FP8 KV cache.")
+
+        if cache_config.cache_dtype == "fp8_e4m3":
+            cache_config.cache_dtype = "fp8_e5m2"
+            logger.warning(
+                "CPU backend doesn't support fp8_e4m3 KV cache type, "
+                "cast to fp8_e5m2.")
+
+        if (cache_config.cache_dtype != "auto"
+                and model_config.dtype == torch.half):
+            logger.warning("FP8 KV cache on the CPU backend only does not"
+                           " support fp16 for now, cast to bf16.")
+            model_config.dtype = torch.bfloat16
+
         kv_cache_space = envs.VLLM_CPU_KVCACHE_SPACE
 
         if kv_cache_space >= 0:
@@ -85,14 +101,6 @@ class CpuPlatform(Platform):
                 "Invalid environment variable VLLM_CPU_KVCACHE_SPACE"
                 f" {kv_cache_space}, expect a positive integer value.")
 
-        scheduler_config = vllm_config.scheduler_config
-        if ((scheduler_config.chunked_prefill_enabled
-             or cache_config.enable_prefix_caching)
-                and model_config.dtype == torch.half):
-            logger.warning("Chunked-prefill on the CPU backend only does not"
-                           " support fp16 for now, cast to bf16.")
-            model_config.dtype = torch.bfloat16
-
         parallel_config = vllm_config.parallel_config
         if (parallel_config.distributed_executor_backend is not None
                 and parallel_config.distributed_executor_backend != "mp"):

vllm/worker/cpu_worker.py

@@ -53,8 +53,11 @@ class CPUCacheEngine:
 
         if cache_config.cache_dtype == "auto":
             self.dtype = model_config.dtype
+        elif cache_config.cache_dtype in ["fp8", "fp8_e5m2"]:
+            self.dtype = torch.float8_e5m2
         else:
-            self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]
+            raise NotImplementedError(f"Unsupported KV cache type "
+                                      f"{cache_config.cache_dtype}.")
 
         # Get attention backend.
         self.attn_backend = get_attn_backend(