fix(server): fix fp8 weight loading

server/text_generation_server/layers/fp8.py

@@ -115,8 +115,12 @@ class HybridFP8UnquantLoader(WeightsLoader):
         return UnquantizedWeight(w)
 
     def get_multi_weights_col(self, weights: "Weights", prefixes: List[str], dim: int):
-        w = [weights.get_sharded(f"{p}.weight", dim=0) for p in prefixes]
+        # FIXME: Force to_device to false as fp8 weights do not support torch.cat on device yet
-        w = torch.cat(w, dim=dim)
+        w = [
+            weights.get_sharded(f"{p}.weight", dim=0, to_device=False) for p in prefixes
+        ]
+        # Concat then send to the device
+        w = torch.cat(w, dim=dim).to(weights.device)
 
         # FP8 branch
         if w.dtype == torch.float8_e4m3fn:

server/text_generation_server/layers/marlin.py

@@ -504,7 +504,7 @@ class GPTQMarlinFP8Linear(nn.Module):
     def __init__(
         self,
         qweight: torch.Tensor,
-        scale: torch.Tensor,
+        scales: torch.Tensor,
         bias: Optional[torch.Tensor],
     ) -> None:
         super().__init__()
@@ -514,8 +514,11 @@ class GPTQMarlinFP8Linear(nn.Module):
 
         log_once(logger.info, "GPU does not support FP8, using Marlin FP8 kernel")
 
-        scale = scale.to(torch.float16)
+        scales = scales.unsqueeze(0)
-        qweight, scales = repack_fp8_for_marlin(qweight, scale)
+        if scales.shape[1] == 1:
+            out_features, in_features = qweight.shape
+            scales = scales.repeat(1, out_features)
+        qweight, scales = repack_fp8_for_marlin(qweight, scales)
 
         in_features = qweight.shape[0] * MARLIN_TILE_SIZE
         out_features = scales.shape[1]
@@ -530,13 +533,13 @@ class GPTQMarlinFP8Linear(nn.Module):
         )
 
     @classmethod
-    def from_unquant(cls, weight, bias, _dtype):
+    def from_unquant(cls, weight, bias, dtype):
-        qweight, scale = fp8_quantize(weight)
+        qweight, scales = fp8_quantize(weight)
-        return cls(qweight=qweight, scale=scale, bias=bias)
+        return cls(qweight=qweight, scales=scales.to(dtype), bias=bias)
 
     @classmethod
-    def from_fp8(cls, weight, scale, _input_scale, bias, _dtype):
+    def from_fp8(cls, weight, scale, _input_scale, bias, dtype):
-        return cls(qweight=weight, scale=scale, bias=bias)
+        return cls(qweight=weight, scales=scale.to(dtype), bias=bias)
 
     def forward(self, A: torch.Tensor) -> torch.Tensor:
         assert marlin_kernels is not None
@@ -591,7 +594,7 @@ def pack_fp8_as_int32(fp8_tensor: torch.Tensor) -> torch.Tensor:
     return packed
 
 
-def repack_fp8_for_marlin(weight: torch.Tensor, scale: torch.Tensor):
+def repack_fp8_for_marlin(weight: torch.Tensor, scales: torch.Tensor):
     """
     Repack FP8 tensor for GPTQ-Marlin.
     """
@@ -608,7 +611,6 @@ def repack_fp8_for_marlin(weight: torch.Tensor, scale: torch.Tensor):
         qweight, perm, in_features, out_features, 8
     )
 
-    scales = scale.reshape(1, 1).repeat(1, out_features)
     scales = permute_scales(scales)
 
     return repacked, scales
@@ -621,7 +623,7 @@ class MarlinWeight(Weight):
 
     Attributes:
         B (torch.Tensor): int4-quantized weights packed into int32.
-        s (torch.Tensor): float16 scales.
+        s (torch.Tensor): bfloat16/float16 scales.
     """
 
     B: torch.Tensor
@@ -629,7 +631,7 @@ class MarlinWeight(Weight):
 
     def __post_init__(self):
         assert self.B.dtype == torch.int32
-        assert self.s.dtype == torch.float16
+        assert self.s.dtype in [torch.float16, torch.bfloat16]
 
     def get_linear(self, bias: torch.Tensor):
         return MarlinLinear(weight=self, bias=bias)

server/text_generation_server/models/__init__.py

@@ -306,16 +306,41 @@ def get_model(
     max_input_tokens: int,
 ) -> Model:
     global FLASH_ATTENTION
+
+    config_dict, _ = PretrainedConfig.get_config_dict(
+        model_id, revision=revision, trust_remote_code=trust_remote_code
+    )
+    model_type = config_dict.get("model_type", None)
+
+    quantization_config = config_dict.get("quantization_config", None)
+    if quantization_config is not None and quantize is None:
+        method = quantization_config.get("quant_method", None)
+        if method in {"gptq", "awq", "exl2"}:
+            log_master(logger.info, f"Auto selecting quantization method {method}")
+            quantize = method
+        elif method == "fbgemm_fp8":
+            log_master(logger.info, "Auto selecting quantization method fp8")
+            quantize = "fp8"
+        else:
+            log_master(logger.warning, f"Unknown quantization method {method}")
+
     if dtype is None:
         if quantize in ["awq", "exl2", "gptq", "marlin"]:
             # These quantizers only work with float16 params.
             dtype = torch.float16
         elif quantize == "fp8":
-            from text_generation_server.layers.fp8 import FBGEMM_MM_AVAILABLE
+            from text_generation_server.layers.fp8 import FBGEMM_DYN_AVAILABLE
 
-            if FBGEMM_MM_AVAILABLE:
+            if FBGEMM_DYN_AVAILABLE:
                 # fbgemm kernels are fp8xfp8->bf16
                 dtype = torch.bfloat16
+        else:
+            config_dtype = config_dict.get("torch_dtype", None)
+            # Only use the config dtype if its one of TGI's supported dtype
+            if config_dtype == "float16":
+                dtype = torch.float16
+            elif config_dtype == "bfloat16":
+                dtype = torch.bfloat16
             else:
                 # Keep it as default for now and let
                 # every model resolve their own default dtype.
@@ -332,11 +357,6 @@ def get_model(
     else:
         set_speculate(0)
 
-    config_dict, _ = PretrainedConfig.get_config_dict(
-        model_id, revision=revision, trust_remote_code=trust_remote_code
-    )
-    model_type = config_dict.get("model_type", None)
-
     speculator = None
     if "medusa_num_heads" in config_dict:
         medusa_model_id = model_id
@@ -451,14 +471,6 @@ def get_model(
             raise RuntimeError(
                 f"Could not determine model type for {model_id} revision {revision}"
             )
-    quantization_config = config_dict.get("quantization_config", None)
-    if quantization_config is not None and quantize is None:
-        method = quantization_config.get("quant_method", None)
-        if method in {"gptq", "awq", "exl2"}:
-            log_master(logger.info, f"Auto selecting quantization method {method}")
-            quantize = method
-        else:
-            log_master(logger.warning, f"Unknown quantization method {method}")
 
     if quantize == "exl2" and sharded:
         raise RuntimeError(

server/text_generation_server/utils/weights.py

@@ -230,7 +230,9 @@ class Weights:
             tensor = tensor.to(device=self.device)
         return tensor
 
-    def get_partial_sharded(self, tensor_name: str, dim: int, to_dtype=True):
+    def get_partial_sharded(
+        self, tensor_name: str, dim: int, to_device=True, to_dtype=True
+    ):
         filename, tensor_name = self.get_filename(tensor_name)
         f = self._get_handle(filename)
         slice_ = f.get_slice(tensor_name)
@@ -256,10 +258,11 @@ class Weights:
             and to_dtype
         ):
             tensor = tensor.to(dtype=self.dtype)
+        if to_device:
             tensor = tensor.to(device=self.device)
         return tensor
 
-    def get_sharded(self, tensor_name: str, dim: int, to_dtype=True):
+    def get_sharded(self, tensor_name: str, dim: int, to_device=True, to_dtype=True):
         filename, tensor_name = self.get_filename(tensor_name)
         f = self._get_handle(filename)
         slice_ = f.get_slice(tensor_name)
@@ -268,7 +271,9 @@ class Weights:
         assert (
             size % world_size == 0
         ), f"The choosen size {size} is not compatible with sharding on {world_size} shards"
-        return self.get_partial_sharded(tensor_name, dim, to_dtype=to_dtype)
+        return self.get_partial_sharded(
+            tensor_name, dim, to_device=to_device, to_dtype=to_dtype
+        )
 
     def get_packed_sharded(
         self,