[ Misc ] `fbgemm` checkpoints (#6559)

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml

@@ -4,8 +4,8 @@ tasks:
 - name: "gsm8k"
   metrics:
   - name: "exact_match,strict-match"
-    value: 0.769
+    value: 0.752
   - name: "exact_match,flexible-extract"
-    value: 0.769
+    value: 0.754
 limit: 1000
 num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform.yaml

+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform -b auto -l 1000 -f 5 -t 1
+model_name: "nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.753
+  - name: "exact_match,flexible-extract"
+    value: 0.753
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
 done
 
 lm_eval --model vllm \
-  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,distributed_executor_backend="ray",trust_remote_code=true,max_model_len=4096 \
+  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,distributed_executor_backend="ray",trust_remote_code=true,max_model_len=4096 \
   --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
   --batch_size $BATCH_SIZE

vllm/_custom_ops.py

@@ -315,6 +315,8 @@ def scaled_fp8_quant(
     Args:
         input: The input tensor to be quantized to FP8
         scale: Optional scaling factor for the FP8 quantization
+        scale_ub: Optional upper bound for scaling factor in dynamic 
+            per token case
         batch_dim_padding: If specified, pad the first dimension
             of the output to at least this value.
         use_per_token_if_dynamic: Whether to do per_tensor or per_token 

vllm/attention/layer.py

@@ -34,6 +34,7 @@ class Attention(nn.Module):
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         blocksparse_params: Optional[Dict[str, Any]] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         if cache_config is not None:
@@ -56,7 +57,7 @@ class Attention(nn.Module):
         self._k_scale = 1.0
         self._v_scale = 1.0
         quant_method = quant_config.get_quant_method(
-            self) if quant_config else None
+            self, prefix=prefix) if quant_config else None
         if quant_method is not None:
             assert isinstance(quant_method, Fp8KVCacheMethod)
             # TODO (mgoin): kv cache dtype should be specified in the FP8

vllm/config.py

@@ -251,7 +251,7 @@ class ModelConfig:
                     f"supported in ROCm.")
             if (self.quantization
                     not in ("fp8", "marlin", "gptq_marlin_24", "gptq_marlin",
-                            "compressed_tensors")):
+                            "fbgemm_fp8", "compressed_tensors")):
                 logger.warning(
                     "%s quantization is not fully "
                     "optimized yet. The speed can be slower than "

vllm/model_executor/layers/fused_moe/layer.py

@@ -182,7 +182,7 @@ class FusedMoE(torch.nn.Module):
             self.quant_method: Optional[QuantizeMethodBase] = (
                 UnquantizedFusedMoEMethod())
         else:
-            self.quant_method = quant_config.get_quant_method(self)
+            self.quant_method = quant_config.get_quant_method(self, prefix)
         assert self.quant_method is not None
 
         self.quant_method.create_weights(

vllm/model_executor/layers/linear.py

@@ -141,6 +141,7 @@ class LinearBase(torch.nn.Module):
         skip_bias_add: bool = False,
         params_dtype: Optional[torch.dtype] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
 
@@ -155,7 +156,8 @@ class LinearBase(torch.nn.Module):
             self.quant_method: Optional[
                 QuantizeMethodBase] = UnquantizedLinearMethod()
         else:
-            self.quant_method = quant_config.get_quant_method(self)
+            self.quant_method = quant_config.get_quant_method(self,
+                                                              prefix=prefix)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         raise NotImplementedError
@@ -182,9 +184,13 @@ class ReplicatedLinear(LinearBase):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: Optional[str] = None):
-        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
-                         quant_config)
+                 prefix: str = ""):
+        super().__init__(input_size,
+                         output_size,
+                         skip_bias_add,
+                         params_dtype,
+                         quant_config,
+                         prefix=prefix)
 
         # All the linear layer supports quant method.
         assert self.quant_method is not None
@@ -258,9 +264,9 @@ class ColumnParallelLinear(LinearBase):
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
                  output_sizes: Optional[List[int]] = None,
-                 prefix: Optional[str] = None):
+                 prefix: str = ""):
         super().__init__(input_size, output_size, skip_bias_add, params_dtype,
-                         quant_config)
+                         quant_config, prefix)
 
         self.gather_output = gather_output
 
@@ -370,7 +376,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: Optional[str] = None):
+                 prefix: str = ""):
         self.output_sizes = output_sizes
         tp_size = get_tensor_model_parallel_world_size()
         assert all(output_size % tp_size == 0 for output_size in output_sizes)
@@ -514,7 +520,7 @@ class QKVParallelLinear(ColumnParallelLinear):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: Optional[str] = None):
+                 prefix: str = ""):
         self.hidden_size = hidden_size
         self.head_size = head_size
         self.total_num_heads = total_num_heads
@@ -707,9 +713,9 @@ class RowParallelLinear(LinearBase):
                  params_dtype: Optional[torch.dtype] = None,
                  reduce_results: bool = True,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: Optional[str] = None):
+                 prefix: str = ""):
         super().__init__(input_size, output_size, skip_bias_add, params_dtype,
-                         quant_config)
+                         quant_config, prefix)
 
         self.input_is_parallel = input_is_parallel
         self.reduce_results = reduce_results

vllm/model_executor/layers/quantization/__init__.py

@@ -10,6 +10,7 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso
     CompressedTensorsConfig)
 from vllm.model_executor.layers.quantization.deepspeedfp import (
     DeepSpeedFPConfig)
+from vllm.model_executor.layers.quantization.fbgemm_fp8 import FBGEMMFp8Config
 from vllm.model_executor.layers.quantization.fp8 import Fp8Config
 from vllm.model_executor.layers.quantization.gptq import GPTQConfig
 from vllm.model_executor.layers.quantization.gptq_marlin import (
@@ -24,6 +25,7 @@ QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
     "awq": AWQConfig,
     "deepspeedfp": DeepSpeedFPConfig,
     "fp8": Fp8Config,
+    "fbgemm_fp8": FBGEMMFp8Config,
     # The order of gptq methods is important for config.py iteration over
     # override_quantization_method(..)
     "marlin": MarlinConfig,

vllm/model_executor/layers/quantization/aqlm.py

@@ -207,8 +207,8 @@ class AQLMConfig(QuantizationConfig):
         return cls(in_group_size, nbits_per_codebook, num_code_books,
                    out_group_size)
 
-    def get_quant_method(
-            self, layer: torch.nn.Module) -> Optional["AQLMLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["AQLMLinearMethod"]:
         if isinstance(layer, LinearBase):
             return AQLMLinearMethod(self)
         return None

vllm/model_executor/layers/quantization/awq.py

@@ -63,8 +63,8 @@ class AWQConfig(QuantizationConfig):
         zero_point = cls.get_from_keys(config, ["zero_point"])
         return cls(weight_bits, group_size, zero_point)
 
-    def get_quant_method(
-            self, layer: torch.nn.Module) -> Optional["AWQLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["AWQLinearMethod"]:
         if isinstance(layer, LinearBase):
             return AWQLinearMethod(self)
         return None

vllm/model_executor/layers/quantization/base_config.py

@@ -97,12 +97,13 @@ class QuantizationConfig(ABC):
             return default
 
     @abstractmethod
-    def get_quant_method(
-            self, layer: torch.nn.Module) -> Optional[QuantizeMethodBase]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional[QuantizeMethodBase]:
         """Get the quantize method to use for the quantized layer.
         
         Args:
             layer: The layer for the quant method.
+            prefix: The full name of the layer in the state dict
         Returns:
             The quantize method. None if the given layer doesn't support quant
             method.

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -60,9 +60,8 @@ class BitsAndBytesConfig(QuantizationConfig):
             target_modules = cls.get_from_keys(config, ["target_modules"])
         return cls(adapter_name, target_modules)
 
-    def get_quant_method(
-            self,
-            layer: torch.nn.Module) -> Optional["BitsAndBytesLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["BitsAndBytesLinearMethod"]:
         if isinstance(layer, LinearBase):
             return BitsAndBytesLinearMethod(self)
         return None

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -44,8 +44,12 @@ class CompressedTensorsConfig(QuantizationConfig):
     def get_name(self) -> str:
         return "compressed_tensors"
 
+    # TODO (@robertgshaw2-neuralmagic): do layer skipping though here
+    # rather than though create_weights to match other methods
     def get_quant_method(
-            self, layer: torch.nn.Module
+        self,
+        layer: torch.nn.Module,
+        prefix: str,
     ) -> Optional["CompressedTensorsLinearMethod"]:
         if isinstance(layer, LinearBase):
             return CompressedTensorsLinearMethod(self)

vllm/model_executor/layers/quantization/deepspeedfp.py

@@ -69,9 +69,8 @@ class DeepSpeedFPConfig(QuantizationConfig):
             "quantize_config.json",
         ]
 
-    def get_quant_method(
-            self,
-            layer: torch.nn.Module) -> Optional["DeepSpeedFPLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["DeepSpeedFPLinearMethod"]:
         if isinstance(layer, LinearBase):
             return DeepSpeedFPLinearMethod(self)
         return None

vllm/model_executor/layers/quantization/fbgemm_fp8.py

+from typing import Any, Dict, List, Optional
+
+import torch
+from torch.nn import Module
+from torch.nn.parameter import Parameter
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
+                                               UnquantizedLinearMethod)
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig, QuantizeMethodBase)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    apply_fp8_linear, create_per_channel_scale_param)
+from vllm.model_executor.utils import set_weight_attrs
+
+logger = init_logger(__name__)
+
+# Note: this is a hack. We should update each model to register the
+# stacked params and get it from there instead in a future PR.
+# fused_name: List[shard_name]
+_FUSED_LAYER_NAME_MAPPING = {
+    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+    "gate_up_proj": ["gate_proj", "up_proj"]
+}
+
+
+class FBGEMMFp8Config(QuantizationConfig):
+    """Config class for FBGEMM Fp8."""
+
+    def __init__(self, ignore_list: List[str], input_scale_ub: float):
+        self.ignore_list = ignore_list
+        self.input_scale_ub = input_scale_ub
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "fbgemm_fp8"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.bfloat16, torch.float16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 89
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "FBGEMMFp8Config":
+        ignore_list = cls.get_from_keys(config, ["modules_to_not_convert"])
+        input_scale_ub = cls.get_from_keys(config, ["activation_scale_ub"])
+        return cls(ignore_list=ignore_list, input_scale_ub=input_scale_ub)
+
+    def _is_layer_skipped(self, prefix: str) -> bool:
+        # prefix: model.layers.0.self_attn.q_proj
+        # proj_name: q_proj
+        proj_name = prefix.split(".")[-1]
+        if proj_name in _FUSED_LAYER_NAME_MAPPING:
+            shard_prefixes = [
+                prefix.replace(proj_name, shard_proj_name)
+                for shard_proj_name in _FUSED_LAYER_NAME_MAPPING[proj_name]
+            ]
+
+            is_skipped = None
+            for shard_prefix in shard_prefixes:
+                is_shard_skipped = shard_prefix in self.ignore_list
+
+                if is_skipped is None:
+                    is_skipped = is_shard_skipped
+                elif is_shard_skipped != is_skipped:
+                    raise ValueError(
+                        f"Detected some but not all shards of {prefix} "
+                        "are quantized. All shards of fused layers "
+                        "to have the same precision.")
+        else:
+            is_skipped = prefix in self.ignore_list
+
+        assert is_skipped is not None
+        return is_skipped
+
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["QuantizeMethodBase"]:
+        if isinstance(layer, LinearBase):
+            if self._is_layer_skipped(prefix):
+                return UnquantizedLinearMethod()
+            return FBGEMMFp8LinearMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class FBGEMMFp8LinearMethod(LinearMethodBase):
+
+    def __init__(self, quant_config: FBGEMMFp8Config):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        del input_size, output_size
+        output_size_per_partition = sum(output_partition_sizes)
+
+        layer.logical_widths = output_partition_sizes
+
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.orig_dtype = params_dtype
+
+        # WEIGHT
+        weight = Parameter(torch.empty(output_size_per_partition,
+                                       input_size_per_partition,
+                                       dtype=torch.float8_e4m3fn),
+                           requires_grad=False)
+        layer.register_parameter("weight", weight)
+        set_weight_attrs(weight, {
+            "input_dim": 1,
+            "output_dim": 0,
+            **extra_weight_attrs,
+        })
+
+        # WEIGHT SCALE
+        weight_scale = create_per_channel_scale_param(output_partition_sizes,
+                                                      **extra_weight_attrs)
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE UPPER BOUND
+        input_scale_ub = torch.nn.Parameter(torch.tensor(
+            (self.quant_config.input_scale_ub), dtype=torch.float32),
+                                            requires_grad=False)
+        layer.input_scale_ub = input_scale_ub
+
+    def process_weights_after_loading(self, layer: Module) -> None:
+        weight = layer.weight
+        layer.weight = Parameter(weight.t(), requires_grad=False)
+
+    def apply(self,
+              layer: torch.nn.Module,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
+        return apply_fp8_linear(input=x,
+                                weight=layer.weight,
+                                weight_scale=layer.weight_scale,
+                                input_scale=None,
+                                input_scale_ub=layer.input_scale_ub,
+                                bias=bias,
+                                cutlass_fp8_supported=True,
+                                use_per_token_if_dynamic=True)

vllm/model_executor/layers/quantization/fp8.py

@@ -66,8 +66,8 @@ class Fp8Config(QuantizationConfig):
         return cls(is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized,
                    activation_scheme=activation_scheme)
 
-    def get_quant_method(
-            self, layer: torch.nn.Module) -> Optional["QuantizeMethodBase"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["QuantizeMethodBase"]:
         from vllm.attention.layer import Attention  # Avoid circular import
 
         if isinstance(layer, LinearBase):

vllm/model_executor/layers/quantization/gptq.py

@@ -69,8 +69,8 @@ class GPTQConfig(QuantizationConfig):
                                                  default=False)
         return cls(weight_bits, group_size, desc_act, lm_head_quantized)
 
-    def get_quant_method(
-            self, layer: torch.nn.Module) -> Optional["GPTQLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["GPTQLinearMethod"]:
         if (isinstance(layer, LinearBase) or
             (isinstance(layer, ParallelLMHead) and self.lm_head_quantized)):
             return GPTQLinearMethod(self)

vllm/model_executor/layers/quantization/gptq_marlin.py

@@ -94,9 +94,8 @@ class GPTQMarlinConfig(QuantizationConfig):
                         " faster inference")
         return None
 
-    def get_quant_method(
-            self,
-            layer: torch.nn.Module) -> Optional["GPTQMarlinLinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["GPTQMarlinLinearMethod"]:
         if (isinstance(layer, LinearBase) or
             (isinstance(layer, ParallelLMHead) and self.lm_head_quantized)):
             return GPTQMarlinLinearMethod(self)

vllm/model_executor/layers/quantization/gptq_marlin_24.py

@@ -109,9 +109,8 @@ class GPTQMarlin24Config(QuantizationConfig):
 
         return None
 
-    def get_quant_method(
-            self,
-            layer: torch.nn.Module) -> Optional["GPTQMarlin24LinearMethod"]:
+    def get_quant_method(self, layer: torch.nn.Module,
+                         prefix: str) -> Optional["GPTQMarlin24LinearMethod"]:
         if isinstance(layer, LinearBase):
             return GPTQMarlin24LinearMethod(self)
         return None