Merge pull request #94 from andrea-fasoli/int8_llm_tp

feat: INT8 LLM TP>1 enablement

Author: chichun-charlie-liu

fms_mo/aiu_addons/i8i8/i8i8_aiu_linear.py

@@ -188,6 +188,27 @@ def forward(self, x):
             self.smoothquant,
         )
 
+    def re_register_qdata(self) -> None:
+        """Remove existing self.qdata tensor and register it again as a buffer.
+        This method is used during TP, after other quantization metadata have been
+        updated.
+        """
+
+        del self.qdata
+        self.register_buffer(
+            "qdata",
+            torch.cat(
+                (
+                    self.w_clip_val,
+                    self.w_clip_valn,
+                    self.a_clip_val,
+                    self.a_clip_valn,
+                    self.zero_shift,
+                    self.smoothquant_scale,
+                )
+            ),
+        )
+
     def __repr__(self) -> str:
         return (
             f"{self.__class__.__name__}"
@@ -222,7 +243,7 @@ def get_int8_aiu_linear(
     # Preprocess linear_config if its linear_type field is a callable
     # (which would not initialize correctly the dataclass parameters).
     # We don't want to alter the original linear_config though.
-    linear_config_for_dataclass: Optional[dict[Union[str, Callable], Any]] = None
+    linear_config_for_dataclass = None
     if callable(linear_config["linear_type"]):
         linear_config_for_dataclass = update_from_partial(linear_config)
         linear_config_for_dataclass["linear_type"] = linear_type
@@ -240,6 +261,36 @@ def get_int8_aiu_linear(
     return linear
 
 
+def is_w_clip_per_channel(
+    w_clip: torch.Tensor,
+) -> bool:
+    """Determine whether the weight clip value in use for INT8 quantization of the
+    provided linear module is:
+    - per-tensor (1 element, 1-dim tensor), or
+    - per-channel (out_feat elements, 1-dim tensor).
+    """
+
+    if w_clip.dim() != 1:
+        raise ValueError(
+            f"TP error: weight clip value dimensions {str(list(w_clip.size()))} are "
+            "incompatible with expected per-tensor or per-channel quantization."
+        )
+    return w_clip.numel() > 1
+
+
+def is_smoothquant_enabled(
+    smoothquant_scale: torch.Tensor,
+) -> bool:
+    """Determine whether smoothquant is enabled on a module."""
+
+    if smoothquant_scale.dim() != 1:
+        raise ValueError(
+            "TP error: smoothquant_scale array should always be 1-dimensional but "
+            f"has size {str(list(smoothquant_scale.size()))}"
+        )
+    return smoothquant_scale.numel() > 1
+
+
 def shard_int8_aiu_linear(
     tensor_values: dict[str, torch.Tensor],
     tp_module: TPModule,
@@ -259,49 +310,73 @@ def shard_int8_aiu_linear(
               | bias     |   0   |  -  |
               | others*  |   N   |  -  |
 
-    Other quantization parameters: w_clip_val, w_clip_valn,
-    a_clip_val, a_clip_valn, zero_shift, smoothquant_scale
-    No sharding on all these parameters, except w_clip_val and w_clip_valn when
-    per-channel quantization is used
+    Other quantization parameters: w_clip_val, w_clip_valn, a_clip_val, a_clip_valn,
+    zero_shift, smoothquant_scale
+
+    No sharding on any of these parameters (they are CLONED on each rank), with the
+    exception of:
+    - w_clip_val and w_clip_valn, only column-sharding and only when per-channel
+      quantization is used
+    - smoothquant_scale, only row-sharding and only if smoothquant in use
+
+    These parameters are 1-dimensional, so if sharding is needed, it is always applied
+    on dim=0.
     """
+
     param_sharding_info: dict[str, dict[str, LinearParameterShardingInfo]] = {}
+    w_clip_linear_param = None
     for module_name, module_info in module_sharding_info.items():
-        int8_aiu_mod = module_info.linear_module
+        int8_aiu_module = module_info.linear_module
+
+        # check every module if per-channel in use (sharding depends on module)
+        if is_w_clip_per_channel(module_info.linear_module.w_clip_val):
+            w_clip_linear_param = LinearParameterShardingInfo(
+                0,
+                ShardType.SHARD if module_info.sharding_dim == 0 else ShardType.CLONE,
+            )
+        else:
+            w_clip_linear_param = LinearParameterShardingInfo(0, ShardType.CLONE)
+
+        # check for every linear module if smoothquant is enabled
+        if is_smoothquant_enabled(module_info.linear_module.smoothquant_scale):
+            smoothquant_linear_param = LinearParameterShardingInfo(
+                0, ShardType.SHARD if module_info.sharding_dim == 1 else ShardType.CLONE
+            )
+        else:
+            smoothquant_linear_param = LinearParameterShardingInfo(0, ShardType.CLONE)
+
         params: dict[str, LinearParameterShardingInfo] = {
             "weight": LinearParameterShardingInfo(
                 module_info.sharding_dim, ShardType.SHARD
             ),
-            # FIXME: with per-channel W, clips need to be sharded
-            # but if per-tensor w, there should be no sharding
-            # HOW CAN WE DISCRIMINATE THE TWO CASES?
-            "w_clip_val": LinearParameterShardingInfo(0, ShardType.CLONE),
-            "w_clip_valn": LinearParameterShardingInfo(0, ShardType.CLONE),
-            # "w_clip_val": LinearParameterShardingInfo(
-            #     module_info.sharding_dim,
-            #     ShardType.SHARD if module_info.sharding_dim == 0 else ShardType.RANK0,
-            # ),
-            # "w_clip_valn": LinearParameterShardingInfo(
-            #     module_info.sharding_dim,
-            #     ShardType.SHARD if module_info.sharding_dim == 0 else ShardType.RANK0,
-            # ),
+            "w_clip_val": w_clip_linear_param,
+            "w_clip_valn": w_clip_linear_param,
             "a_clip_val": LinearParameterShardingInfo(0, ShardType.CLONE),
             "a_clip_valn": LinearParameterShardingInfo(0, ShardType.CLONE),
             "zero_shift": LinearParameterShardingInfo(0, ShardType.CLONE),
-            "smooqthquant_scale": LinearParameterShardingInfo(0, ShardType.CLONE),
+            "smoothquant_scale": smoothquant_linear_param,
         }
-        if int8_aiu_mod.bias is not None:
+        if int8_aiu_module.bias is not None and int8_aiu_module.bias.numel() > 1:
             params["bias"] = LinearParameterShardingInfo(
-                module_info.sharding_dim,
+                0,
                 ShardType.SHARD if module_info.sharding_dim == 0 else ShardType.RANK0,
             )
         param_sharding_info[module_name] = params
 
+    # trim qdata from dictionary of tensors to be copied on sharded modules.
+    # if not trimmed, qdata wouldn't be copied but the keys would be marked as unused
+    tensor_values = {k: v for k, v in tensor_values.items() if "qdata" not in k}
+
     unused_keys = shard_base_linear(
         tensor_values, tp_module, module_sharding_info, param_sharding_info
     )
 
-    raise NotImplementedError("TP not yet supported for INT8. Work in progress")
-    # return unused_keys
+    # qdata contains all quantization metadata to pass to the AIU and needs to be
+    # updated post-sharding, after metadata tensor have changed
+    for module_name, module_info in module_sharding_info.items():
+        module_info.linear_module.re_register_qdata()
+
+    return unused_keys
 
 
 register_linear_type_to_module_map(
@@ -320,4 +395,6 @@ def shard_int8_aiu_linear(
         use_smoothquant=True,
     ),
 )
+
+# int8 linear with and w/o smoothquant share a common sharding map
 register_linear_type_to_sharding_map("int8_aiu", shard_int8_aiu_linear)