Add support for FP8 KV cache scales

Since FP8 only has limited dynamic range, we can scale keys/values before storing them into the cache (and unscale them in attention). To avoid rescaling the cache as the absmax values change, good scales are usually determined per layer using calibration calibration data and stored in the checkpoint. This change adds support for for using key-value scales and loading them from checkpoints in the two most common formats: - Separate per-layer `k_scale` and `v_scale` scalars. - Per-layer `kv_scale` scalar (older format). Currently, scales are only used with an `float8_e4m3fn` cache.
huggingface · Oct 21, 2024 · a4cb3d3 · a4cb3d3
1 parent 5e0fb46
commit a4cb3d3
Show file tree

Hide file tree

Showing 29 changed files with 332 additions and 49 deletions.
diff --git a/flake.lock b/flake.lock
diff --git a/flake.nix b/flake.nix
@@ -5,7 +5,7 @@
       inputs.nixpkgs.follows = "tgi-nix/nixpkgs";
     };
     nix-filter.url = "github:numtide/nix-filter";
-    tgi-nix.url = "github:huggingface/text-generation-inference-nix";
+    tgi-nix.url = "github:huggingface/text-generation-inference-nix/marlin-kernels-0.3.0";
     nixpkgs.follows = "tgi-nix/nixpkgs";
     flake-utils.url = "github:numtide/flake-utils";
     rust-overlay = {

diff --git a/server/poetry.lock b/server/poetry.lock
diff --git a/server/pyproject.toml b/server/pyproject.toml
@@ -41,10 +41,10 @@ py-cpuinfo = "^9.0.0"
 numpy = "^1.26"
 
 marlin-kernels = [
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.2.0/marlin_kernels-0.2.0+cu123torch2.4-cp39-cp39-linux_x86_64.whl", python = "~3.9", optional = true },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.2.0/marlin_kernels-0.2.0+cu123torch2.4-cp310-cp310-linux_x86_64.whl", python = "~3.10", optional = true },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.2.0/marlin_kernels-0.2.0+cu123torch2.4-cp311-cp311-linux_x86_64.whl", python = "~3.11", optional = true },
-  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.2.0/marlin_kernels-0.2.0+cu123torch2.4-cp312-cp312-linux_x86_64.whl", python = "~3.12", optional = true },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.0/marlin_kernels-0.3.0+cu123torch2.4-cp39-cp39-linux_x86_64.whl", python = "~3.9", optional = true },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.0/marlin_kernels-0.3.0+cu123torch2.4-cp310-cp310-linux_x86_64.whl", python = "~3.10", optional = true },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.0/marlin_kernels-0.3.0+cu123torch2.4-cp311-cp311-linux_x86_64.whl", python = "~3.11", optional = true },
+  { url = "https://github.com/danieldk/marlin-kernels/releases/download/v0.3.0/marlin_kernels-0.3.0+cu123torch2.4-cp312-cp312-linux_x86_64.whl", python = "~3.12", optional = true },
 ]
 moe-kernels = [
   { url = "https://github.com/danieldk/moe-kernels/releases/download/v0.6.0/moe_kernels-0.6.0+cu123torch2.4-cp39-cp39-linux_x86_64.whl", python = "~3.9", optional = true },

diff --git a/server/text_generation_server/layers/attention/__init__.py b/server/text_generation_server/layers/attention/__init__.py
@@ -28,10 +28,11 @@
     raise ImportError(f"System {SYSTEM} doesn't support flash/paged attention")
 
 # KVCache needs `reshape_and_cache`, so ensure that it is defined already.
-from .kv_cache import KVCache
+from .kv_cache import KVCache, get_kv_scales
 
 __all__ = [
     "attention",
+    "get_kv_scales",
     "paged_attention",
     "SUPPORTS_WINDOWING",
     "KVCache",

diff --git a/server/text_generation_server/layers/attention/cuda.py b/server/text_generation_server/layers/attention/cuda.py
@@ -1,5 +1,5 @@
 import torch
-from text_generation_server.layers.attention.kv_cache import KVCache
+from text_generation_server.layers.attention.kv_cache import KVCache, KVScales
 from text_generation_server.utils.import_utils import SYSTEM
 from text_generation_server.models.globals import (
     ATTENTION,
@@ -8,6 +8,7 @@
 from text_generation_server.layers.attention import Seqlen
 from typing import Optional
 
+
 major, minor = torch.cuda.get_device_capability()
 is_sm75 = major == 7 and minor == 5
 _PARTITION_SIZE = 512
@@ -21,6 +22,8 @@ def paged_attention(
     block_tables: torch.Tensor,
     seqlen: Seqlen,
     max_s: int,
+    *,
+    kv_scales: KVScales,
     softcap: Optional[float] = None,
 ):
     # Adapted from: https://github.com/vllm-project/vllm/blob/f8a1e39fae05ca610be8d5a78be9d40f5274e5fc/vllm/model_executor/layers/attention.py
@@ -46,6 +49,8 @@ def paged_attention(
     num_seqs, num_heads, head_size = query.shape
     max_num_partitions = (max_s + _PARTITION_SIZE - 1) // _PARTITION_SIZE
 
+    can_scale = kv_cache.can_scale(kv_scales)
+
     # NOTE(woosuk): We use a simple heuristic to decide whether to use
     # PagedAttention V1 or V2. If the number of partitions is 1, we use
     # V1 to avoid the overhead of reduction. Also, if the number of
@@ -59,6 +64,8 @@ def paged_attention(
             paged_kv_cache=(kv_cache.key, kv_cache.value),
             logits_soft_cap=softcap,
             sm_scale=softmax_scale,
+            k_scale=kv_scales.key_scale_cpu if can_scale else 1.0,
+            v_scale=kv_scales.value_scale_cpu if can_scale else 1.0,
         )
     elif ATTENTION == "flashdecoding":
         max_q = 1
@@ -204,13 +211,16 @@ def attention(
     key: torch.Tensor,
     value: torch.Tensor,
     kv_cache: KVCache,
+    kv_scales: KVScales,
     seqlen: Seqlen,
     block_tables: torch.Tensor,
     softmax_scale: float,
     window_size_left: int = -1,
     causal: bool = True,
     softcap: Optional[float] = None,
 ):
+    can_scale = kv_cache.can_scale(kv_scales)
+
     if ATTENTION == "flashinfer":
         from text_generation_server.layers.attention.flashinfer import (
             prefill_with_paged_kv_state,
@@ -226,6 +236,8 @@ def attention(
             logits_soft_cap=softcap,
             sm_scale=softmax_scale,
             window_left=window_size_left,
+            k_scale=kv_scales.key_scale_cpu if can_scale else 1.0,
+            v_scale=kv_scales.value_scale_cpu if can_scale else 1.0,
         )
 
     # If we are using flashdecoding or paged, we always use flash-attn for

diff --git a/server/text_generation_server/layers/attention/flashinfer.py b/server/text_generation_server/layers/attention/flashinfer.py
@@ -204,6 +204,7 @@ def use_decode_state(
     num_kv_heads: int,
     head_size: int,
     page_size: int,
+    kv_cache_dtype: torch.dtype,
     dtype: torch.dtype,
     window_left: int,
 ):
@@ -240,7 +241,7 @@ def use_decode_state(
             num_kv_heads=num_kv_heads,
             head_dim=head_size,
             page_size=page_size,
-            data_type=dtype,
+            data_type=kv_cache_dtype,
             q_data_type=dtype,
             window_left=window_left,
         )

diff --git a/server/text_generation_server/layers/attention/ipex.py b/server/text_generation_server/layers/attention/ipex.py
@@ -1,6 +1,6 @@
 import intel_extension_for_pytorch as ipex
 import torch
-from text_generation_server.layers.attention.kv_cache import KVCache
+from text_generation_server.layers.attention.kv_cache import KVCache, KVScales
 from text_generation_server.models.flash_causal_lm import BLOCK_SIZE
 from text_generation_server.layers.attention import Seqlen
 from typing import Optional
@@ -14,6 +14,7 @@ def attention(
     key: torch.Tensor,
     value: torch.Tensor,
     kv_cache: KVCache,
+    kv_scales: KVScales,
     seqlen: Seqlen,
     block_tables: torch.Tensor,
     softmax_scale: float,
@@ -55,6 +56,8 @@ def paged_attention(
     block_tables: torch.Tensor,
     seqlen: Seqlen,
     max_s: int,
+    *,
+    kv_scales: KVScales,
     softcap: Optional[float] = None,
 ):
     if softcap is not None:

diff --git a/server/text_generation_server/layers/attention/kv_cache.py b/server/text_generation_server/layers/attention/kv_cache.py
@@ -1,8 +1,38 @@
 from typing import Tuple
+from dataclasses import dataclass, field
 
+from loguru import logger
 import torch
+
+from text_generation_server.layers.fp8 import fp8_quantize
 from text_generation_server.models.globals import ATTENTION, BLOCK_SIZE
 from text_generation_server.utils.import_utils import SYSTEM
+from text_generation_server.utils.log import log_once
+from text_generation_server.utils.weights import Weights
+
+
+@dataclass
+class KVScales:
+    """
+    Key-value scales for FP8 KV cache.
+
+    This data class stores key and value scales both as a GPU tensor and
+    as a GPU float. This inconvenience is necessary because some functions
+    (e.g. scaling kernels) take scales as a GPU tensor, whereas others
+    (e.g. flashinfer) take scales as a CPU scalar.
+    """
+
+    key_scale: torch.Tensor
+    value_scale: torch.Tensor
+    key_scale_cpu: float = field(init=False)
+    value_scale_cpu: float = field(init=False)
+
+    def __post_init__(self):
+        if self.key_scale.numel() != 1 or self.value_scale.numel() != 1:
+            raise ValueError("Key and value scales must be scalar tensors.")
+
+        self.key_scale_cpu = self.key_scale.item()
+        self.value_scale_cpu = self.value_scale.item()
 
 
 class KVCache:
@@ -76,6 +106,29 @@ def __init__(
                 ),
             )
 
+    def can_scale(self, kv_scales: KVScales) -> bool:
+        """Check if the cache can be scaled by the given scales."""
+        if kv_scales.key_scale_cpu == 1.0 and kv_scales.value_scale_cpu == 1.0:
+            return False
+        elif self.dtype == torch.float8_e4m3fn and SYSTEM == "cuda":
+            log_once(
+                logger.info,
+                "Using FP8 KV cache scales",
+            )
+            return True
+        else:
+            # We have scales, but not the correct FP8 cache type, so warn once.
+            log_once(
+                logger.info,
+                "Ignoring FP8 KV cache scales, only float8_e4m3fn KV cache on CUDA is supported",
+            )
+            return False
+
+    @property
+    def dtype(self):
+        """Get the data type of the cache."""
+        return self.kv_cache[0].dtype
+
     @property
     def key(self):
         """Get the key cache."""
@@ -94,17 +147,33 @@ def store(
         key: torch.Tensor,
         value: torch.Tensor,
         slots: torch.Tensor,
+        kv_scales: KVScales,
     ):
         """Store the key and value at the given slots."""
 
         key_cache = self.kv_cache[0]
         value_cache = self.kv_cache[1]
 
+        if self.can_scale(kv_scales):
+            if kv_scales.key_scale_cpu != 1.0:
+                key = fp8_quantize(
+                    key.float(),
+                    scale=kv_scales.key_scale,
+                    qdtype=self.dtype,
+                    scalar=True,
+                )[0]
+            if kv_scales.value_scale_cpu != 1.0:
+                value = fp8_quantize(
+                    value.float(),
+                    scale=kv_scales.value_scale,
+                    qdtype=self.dtype,
+                    scalar=True,
+                )[0]
+
         if ATTENTION in {"flashdecoding", "flashinfer"}:
-            # TODO: add scale
             key = key.to(key_cache.dtype)
             value = value.to(value_cache.dtype)
-            if key_cache.dtype in {torch.float8_e5m2, torch.float8_e4m3fn}:
+            if key_cache.dtype in {torch.float8_e4m3fn, torch.float8_e5m2}:
                 # Torch index_put does not support float8_{e5m2,e4m3fn} yet, so
                 # put as raw data instead.
                 key_cache = key_cache.view(torch.uint8)
@@ -151,5 +220,23 @@ def paged_reshape_and_cache(
         )
     else:
         raise NotImplementedError(
-            f"Cannot reshape and cache for paged attention, system '{SYSTEM}' not supportedattention"
+            f"Cannot reshape and cache for paged attention, system '{SYSTEM}' not supported"
         )
+
+
+def get_kv_scales(weights: Weights, prefix: str) -> KVScales:
+    """Load KV cache scales."""
+
+    key_scale = torch.tensor(1.0, dtype=torch.float32, device=weights.device)
+    value_scale = key_scale
+    if weights.has_tensor(f"{prefix}.k_scale") and weights.has_tensor(
+        f"{prefix}.v_scale"
+    ):
+        key_scale = weights.get_tensor(f"{prefix}.k_scale", to_dtype=False).float()
+        value_scale = weights.get_tensor(f"{prefix}.v_scale", to_dtype=False).float()
+    elif weights.has_tensor(f"{prefix}.kv_scale"):
+        # Fall back to older more coarse-grained scale when available.
+        key_scale = weights.get_tensor(f"{prefix}.kv_scale").float()
+        value_scale = key_scale
+
+    return KVScales(key_scale=key_scale, value_scale=value_scale)