Support calibrating kv cache scales (#17)

* Support calibrating kv cache scales

* Add comment

* Fix weight name

* Add Qwen test

* Fix kv cache test count

* Add fixed target sizes

* Fix proj linear count

* Switch from output_scale to kv_scale

* Add example

Author: mgoin

auto_fp8/config.py

@@ -1,4 +1,4 @@
-from typing import List
+from typing import List, Optional, Tuple
 
 
 class BaseQuantizeConfig:
@@ -17,13 +17,17 @@ class BaseQuantizeConfig:
             regex style matching i.e. re.search(), for each Linear layer.
             By default, "re:.*lm_head" is included to ignore the embedding
             Linear layer usually at the end of decoder LLMs
+        kv_cache_quant_targets: Tuple of Linear module names to target for
+            calibration of the output scales for KV cache quantization.
+            Usually, these should be `("k_proj", "v_proj")`.
     """
 
     def __init__(
         self,
         quant_method: str = "fp8",
         activation_scheme: str = "static",
-        ignore_patterns: List[str] = [],
+        ignore_patterns: List[str] = ["re:.*lm_head"],
+        kv_cache_quant_targets: Optional[Tuple[str]] = None,
     ):
         if quant_method != "fp8":
             raise ValueError("Only FP8 quantization is supported.")
@@ -34,4 +38,5 @@ def __init__(
         self.quant_method = quant_method
         self.activation_scheme = activation_scheme
         self.ignore_patterns = ignore_patterns
+        self.kv_cache_quant_targets = kv_cache_quant_targets
         self.ignored_layers = []

auto_fp8/modeling.py

@@ -1,5 +1,5 @@
 import re
-from typing import List
+from typing import List, Optional, Tuple
 
 import torch
 from transformers import AutoModelForCausalLM
@@ -27,6 +27,16 @@ def __init__(
             self.model, quantize_config.ignore_patterns
         )
 
+        if quantize_config.kv_cache_quant_targets:
+            kv_cache_quant_layers = get_kv_cache_quant_layers(
+                self.model, quantize_config.kv_cache_quant_targets
+            )
+            if len(kv_cache_quant_layers) == 0:
+                raise ValueError(
+                    f"Could not find any kv cache layers using kv_cache_quant_targets={quantize_config.kv_cache_quant_targets}, please fix your argument."
+                )
+            quantize_config.kv_cache_quant_layers = kv_cache_quant_layers
+
         self.quantize_config = quantize_config
 
     @classmethod
@@ -97,26 +107,28 @@ def skip(*args, **kwargs):
 
         return cls(model, quantize_config)
 
-    def quantize(self, calibration_tokens):
-        def _prepare_calibration_data(calibration_tokens):
-            if hasattr(calibration_tokens, "input_ids"):
-                return calibration_tokens.input_ids
-            return calibration_tokens
+    def quantize(self, calibration_tokens: Optional[torch.Tensor] = None):
 
         # Always quantize the weights as they do not require calibration data
         quantize_weights(self.model, self.quantize_config)
 
         if self.quantize_config.activation_scheme == "static":
+            assert (
+                calibration_tokens is not None
+            ), "Calibration tokens required for activation quantization"
+
+
+            def _prepare_calibration_data(calibration_tokens):
+                if hasattr(calibration_tokens, "input_ids"):
+                    return calibration_tokens.input_ids
+                return calibration_tokens
+
             quantize_activations(
                 self.model,
                 self.quantize_config,
                 _prepare_calibration_data(calibration_tokens),
             )
 
-            # import copy
-            # for layer in self.model.model.layers:
-            #     layer.self_attn.kv_scale = copy.deepcopy(layer.self_attn.k_proj.input_scale)
-
     def save_quantized(self, save_dir):
         save_quantized_model(
             self.model,
@@ -128,9 +140,6 @@ def save_quantized(self, save_dir):
 def get_layers_to_ignore(model, ignore_patterns) -> List[str]:
     ignored_layers = set()
 
-    # TODO: don't always ignore lm_head
-    ignore_patterns.append("re:.*lm_head")
-
     for name, linear in model.named_modules():
         if not isinstance(linear, torch.nn.Linear):
             continue
@@ -148,3 +157,17 @@ def get_layers_to_ignore(model, ignore_patterns) -> List[str]:
                     ignored_layers.add(name)
 
     return list(ignored_layers)
+
+
+def get_kv_cache_quant_layers(model, kv_cache_quant_targets: Tuple[str]) -> List[str]:
+    kv_cache_quant_layers = []
+
+    for name, linear in model.named_modules():
+        if not isinstance(linear, torch.nn.Linear):
+            continue
+
+        for output_quant_target in kv_cache_quant_targets:
+            if name.endswith(output_quant_target):
+                kv_cache_quant_layers.append(name)
+
+    return kv_cache_quant_layers

auto_fp8/quantize.py

@@ -1,6 +1,6 @@
 import gc
 import re
-from typing import List, Tuple
+from typing import Optional, Tuple
 import copy
 
 import torch
@@ -61,14 +61,22 @@ def per_tensor_quantize(tensor: torch.Tensor) -> Tuple[torch.Tensor, float]:
     return qweight, scale
 
 
+def static_per_tensor_quantize(tensor: torch.Tensor, inv_scale: float) -> torch.Tensor:
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    qweight = (tensor / inv_scale).clamp(min=finfo.min, max=finfo.max)
+    return qweight.to(torch.float8_e4m3fn)
+
+
 def fp8_gemm(A, A_scale, B, B_scale, bias, out_dtype):
     if A.numel() == 0:
         # Deal with empty tensors (triggeted by empty MoE experts)
         return torch.empty(size=(0, B.shape[0]), dtype=out_dtype, device=A.device)
-    
-    native_fp8_support = (
-        torch.cuda.is_available() and torch.cuda.get_device_capability() >= (8, 9)
-    )
+
+    # TODO: Disable native fp8 gemm for now, always just dequantize
+    # native_fp8_support = (
+    #     torch.cuda.is_available() and torch.cuda.get_device_capability() >= (8, 9)
+    # )
+    native_fp8_support = False
     if native_fp8_support:
         need_reshape = A.dim() == 3
         if need_reshape:
@@ -98,25 +106,24 @@ def fp8_gemm(A, A_scale, B, B_scale, bias, out_dtype):
     return output
 
 
-class FP8StaticLinearQuantizer(torch.nn.Module):
+# Class responsible for quantizing weights
+class FP8DynamicLinear(torch.nn.Module):
     def __init__(
-        self, qweight: torch.Tensor, weight_scale: torch.Tensor, bias: torch.Tensor
+        self,
+        weight: torch.Tensor,
+        weight_scale: torch.Tensor,
+        bias: torch.nn.Parameter,
     ):
         super().__init__()
-        self.weight = torch.nn.Parameter(qweight, requires_grad=False)
+        self.weight = torch.nn.Parameter(weight, requires_grad=False)
         self.weight_scale = torch.nn.Parameter(weight_scale, requires_grad=False)
-        self.input_scale = None
         self.bias = bias
 
     def forward(self, x):
-        qinput, x_input_scale = per_tensor_quantize(x)
-        if self.input_scale is None:
-            self.input_scale = torch.nn.Parameter(x_input_scale)
-        elif x_input_scale > self.input_scale:
-            self.input_scale = torch.nn.Parameter(x_input_scale)
+        qinput, x_scale = per_tensor_quantize(x)
         output = fp8_gemm(
             A=qinput,
-            A_scale=self.input_scale,
+            A_scale=x_scale,
             B=self.weight,
             B_scale=self.weight_scale,
             bias=self.bias,
@@ -125,29 +132,29 @@ def forward(self, x):
         return output
 
 
-class FP8StaticLinear(torch.nn.Module):
+# Module responsible for taking already quantized weights, and recording input scales (and possibly output scales) using an activation observer
+class FP8StaticLinearQuantizer(torch.nn.Module):
     def __init__(
         self,
-        qweight: torch.Tensor,
+        weight: torch.Tensor,
         weight_scale: torch.Tensor,
-        bias: torch.Tensor,
-        input_scale: float = 1.0,
+        bias: torch.nn.Parameter,
+        quantize_output: bool = False,
     ):
         super().__init__()
-        self.weight = torch.nn.Parameter(qweight, requires_grad=False)
+        self.weight = torch.nn.Parameter(weight, requires_grad=False)
         self.weight_scale = torch.nn.Parameter(weight_scale, requires_grad=False)
-        self.input_scale = torch.nn.Parameter(input_scale, requires_grad=False)
         self.bias = bias
-
-    def per_tensor_quantize(
-        self, tensor: torch.Tensor, inv_scale: float
-    ) -> torch.Tensor:
-        finfo = torch.finfo(torch.float8_e4m3fn)
-        qweight = (tensor / inv_scale).clamp(min=finfo.min, max=finfo.max)
-        return qweight.to(torch.float8_e4m3fn)
+        self.input_scale = None
+        self.output_scale = None
+        self.quantize_output = quantize_output
 
     def forward(self, x):
-        qinput = self.per_tensor_quantize(x, inv_scale=self.input_scale)
+        qinput, x_input_scale = per_tensor_quantize(x)
+        if self.input_scale is None:
+            self.input_scale = torch.nn.Parameter(x_input_scale, requires_grad=False)
+        elif x_input_scale > self.input_scale:
+            self.input_scale = torch.nn.Parameter(x_input_scale, requires_grad=False)
         output = fp8_gemm(
             A=qinput,
             A_scale=self.input_scale,
@@ -156,26 +163,51 @@ def forward(self, x):
             bias=self.bias,
             out_dtype=x.dtype,
         )
+
+        # Optionally, quantize output and record scale
+        if self.quantize_output:
+            qoutput, output_scale = per_tensor_quantize(output)
+            if self.output_scale is None:
+                self.output_scale = torch.nn.Parameter(output_scale, requires_grad=False)
+            elif output_scale > self.output_scale:
+                self.output_scale = torch.nn.Parameter(output_scale, requires_grad=False)
+            output = qoutput.to(output.dtype) * output_scale
+
         return output
 
 
-class FP8DynamicLinear(torch.nn.Module):
-    def __init__(self, qweight: torch.Tensor, scale: torch.Tensor, bias: torch.Tensor):
+# Module responsible for representing the final checkpoint representation
+class FP8StaticLinear(torch.nn.Module):
+    def __init__(
+        self,
+        weight: torch.nn.Parameter,
+        weight_scale: torch.nn.Parameter,
+        bias: torch.nn.Parameter,
+        input_scale: torch.nn.Parameter,
+        output_scale: Optional[torch.nn.Parameter] = None,
+    ):
         super().__init__()
-        self.weight = torch.nn.Parameter(qweight, requires_grad=False)
-        self.weight_scale = torch.nn.Parameter(scale, requires_grad=False)
+        self.weight = weight
+        self.weight_scale = weight_scale
         self.bias = bias
+        self.input_scale = input_scale
+        self.output_scale = output_scale
 
     def forward(self, x):
-        qinput, x_scale = per_tensor_quantize(x)
+        qinput = static_per_tensor_quantize(x, self.input_scale)
         output = fp8_gemm(
             A=qinput,
-            A_scale=x_scale,
+            A_scale=self.input_scale,
             B=self.weight,
             B_scale=self.weight_scale,
             bias=self.bias,
             out_dtype=x.dtype,
         )
+
+        if self.output_scale:
+            qoutput = static_per_tensor_quantize(output, self.output_scale)
+            output = qoutput.to(output.dtype) * self.output_scale
+
         return output
 
 
@@ -194,7 +226,6 @@ def replace_module(model: AutoModelForCausalLM, name: str, new_module: torch.nn.
 def quantize_weights(
     model: AutoModelForCausalLM,
     quantize_config: BaseQuantizeConfig,
-    ignored_layers: List[str] = [],
 ):
     named_modules = list(model.named_modules())
     for name, linear in tqdm.tqdm(named_modules, desc="Quantizing weights"):
@@ -203,9 +234,11 @@ def quantize_weights(
             or name in quantize_config.ignored_layers
         ):
             continue
-        quant_weight, quant_scale = per_tensor_quantize(linear.weight)
+        quant_weight, weight_scale = per_tensor_quantize(linear.weight)
         bias = copy.deepcopy(linear.bias) if linear.bias is not None else None
-        quant_linear = FP8DynamicLinear(quant_weight, quant_scale, bias)
+        quant_linear = FP8DynamicLinear(
+            weight=quant_weight, weight_scale=weight_scale, bias=bias
+        )
         replace_module(model, name, quant_linear)
         del linear.weight
         del linear.bias
@@ -217,7 +250,6 @@ def quantize_activations(
     model: AutoModelForCausalLM,
     quantize_config: BaseQuantizeConfig,
     calibration_tokens,
-    ignored_layers: List[str] = [],
 ):
     # Replace weight quantizer with a dynamic activation quantizer observer
     for name, dynamic_quant_linear in model.named_modules():
@@ -227,9 +259,13 @@ def quantize_activations(
         ):
             continue
         quantizer = FP8StaticLinearQuantizer(
-            dynamic_quant_linear.weight,
-            dynamic_quant_linear.weight_scale,
-            dynamic_quant_linear.bias,
+            weight=dynamic_quant_linear.weight,
+            weight_scale=dynamic_quant_linear.weight_scale,
+            bias=dynamic_quant_linear.bias,
+            quantize_output=(
+                hasattr(quantize_config, "kv_cache_quant_layers")
+                and name in quantize_config.kv_cache_quant_layers
+            ),
         )
         replace_module(model, name, quantizer)
         del dynamic_quant_linear
@@ -251,21 +287,45 @@ def quantize_activations(
         ):
             continue
         static_proj = FP8StaticLinear(
-            quantizer.weight,
-            quantizer.weight_scale,
-            quantizer.bias,
-            quantizer.input_scale,
+            weight=quantizer.weight,
+            weight_scale=quantizer.weight_scale,
+            bias=quantizer.bias,
+            input_scale=quantizer.input_scale,
+            output_scale=quantizer.output_scale,
         )
         replace_module(model, name, static_proj)
         del quantizer
     cleanup_memory()
 
+    # Post-process step for kv cache scales to take the k/v module
+    # `output_scale` parameters, take the max of them, and store them in
+    # the parent attention module as `kv_scale`
+    # NOTE: if we want to switch to the `output_scale` representation, we can simply remove this block
+    if hasattr(quantize_config, "kv_cache_quant_layers"):
+        # Assumes that list is ordered such that [layer0.k_proj, layer0.v_proj, layer1.k_proj, layer1.v_proj, ...]
+        # so we make a list of tuples [(layer0.k_proj, layer0.v_proj), (layer1.k_proj, layer1.v_proj), ...]
+        kv_proj_pairs = zip(*[iter(quantize_config.kv_cache_quant_layers)]*2)
+        for k_proj_name, v_proj_name in kv_proj_pairs:
+            parent_module_name = ".".join(k_proj_name.split(".")[:-1])
+            assert parent_module_name == ".".join(v_proj_name.split(".")[:-1])
+            parent_module = dict(model.named_modules())[parent_module_name]
+
+            k_proj = dict(model.named_modules())[k_proj_name]
+            v_proj = dict(model.named_modules())[v_proj_name]
+
+            kv_scale = max(k_proj.output_scale, v_proj.output_scale)
+            parent_module.kv_scale = torch.nn.Parameter(kv_scale, requires_grad=False)
+
+            # Remove output_scale from k_proj and v_proj
+            k_proj.output_scale = None
+            v_proj.output_scale = None
+    cleanup_memory()
+
 
 def save_quantized_model(
     model: AutoModelForCausalLM,
     quant_config: BaseQuantizeConfig,
     save_dir: str,
-    ignored_layers: List[str] = [],
 ):
     print(model)
     print(f"Saving the model to {save_dir}")
@@ -276,6 +336,8 @@ def save_quantized_model(
             "ignored_layers": quant_config.ignored_layers,
         }
     }
+    if hasattr(quant_config, "kv_cache_quant_layers"):
+        static_q_dict["quantization_config"]["kv_cache_scheme"] = "static"
     model.config.update(static_q_dict)
     model.save_pretrained(save_dir)
     tokenizer = AutoTokenizer.from_pretrained(model.config._name_or_path)