Feat (graph/equalize): improvements for llm equalization

src/brevitas/graph/base.py

@@ -296,6 +296,15 @@ class FnToModule(CallableToModule):
     def match_node(self, node: Node) -> bool:
         return node.op == 'call_function' and node.target is self.old_callable
 
+    def move_node_args_to_kwargs(self, node: Node):
+        super().move_node_args_to_kwargs(node)
+        # Moving to stateful modules, we remove the 'training' argument if it is passed to the
+        # functional version of the layer since it is not needed anymore
+        kwargs = dict(node.kwargs)
+        if 'training' in kwargs:
+            del kwargs['training']
+        node.kwargs = immutable_dict(kwargs)
+
 
 class MethodToModule(CallableToModule):
 

src/brevitas/graph/equalize.py

@@ -27,7 +27,10 @@
 
 __all__ = ['GraphActivationEqualization', 'LayerwiseActivationEqualization', 'EqualizeGraph']
 
-EPSILON = 1e-9
+# TODO: if we are able to run activation equalization in GPU + float16, we could have two separate
+# epsilon factors for float16 (2e-5) vs float32/bfloat16 (1e-9). At the moment we are tied to one
+# single epsilon for both cases.
+EPSILON = 2e-5
 
 _supported_layers = (
     nn.ConvTranspose1d,
@@ -73,6 +76,8 @@
 
 _batch_norm = (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)
 
+_ignore_ops = (getattr, 'size', 'contiguous')
+
 
 # Required for being hashable
 @dataclass(eq=True, frozen=True)
@@ -398,7 +403,6 @@ def _no_equalize():
     scale_fn = _select_scale_computation_fn(scale_computation_type)
     sink_weights = [transpose(m, axis) for m, axis in sink_axes.items()]
     sinks_range = scale_fn(torch.cat([w.reshape(w.size(0), -1) for w in sink_weights], 1))
-    sinks_range = torch.clamp(sinks_range, EPSILON)
 
     # Determine the srcs_range based on where we are performing activation equalization or
     # weight equalization
@@ -431,10 +435,18 @@ def _no_equalize():
             "Detected source and sink with non compatible shapes, equalization is skipped")
         return _no_equalize()
 
+    # Instead of clipping very low values, which would cause their reciprocal to be very large
+    # thus hindering quantization, we set them to one, which is the no-op equivalent for equalization
+    sinks_range = torch.where((sinks_range < EPSILON) | (srcs_range < EPSILON),
+                              torch.tensor(1., dtype=dtype, device=device),
+                              sinks_range)
+    srcs_range = torch.where((sinks_range < EPSILON) | (srcs_range < EPSILON),
+                             torch.tensor(1., dtype=dtype, device=device),
+                             srcs_range)
+
     srcs_range = torch.pow(srcs_range, alpha)
     sinks_range = torch.pow(sinks_range, 1 - alpha)
     scaling_factors = srcs_range / sinks_range
-    scaling_factors = torch.clamp(scaling_factors, EPSILON)
     inverse_scaling_factors = torch.reciprocal(scaling_factors)
 
     if list_of_act_val is not None and list_of_insert_mul_node_fn is not None:
@@ -455,8 +467,8 @@ def _no_equalize():
                     torch.reshape(inverse_scaling_factors, src_broadcast_size)),
                 attr='weight')
     for module, axis in sink_axes.items():
-        src_broadcast_size = [1] * module.weight.ndim
-        src_broadcast_size[axis] = module.weight.size(axis)
+        sink_broadcast_size = [1] * module.weight.ndim
+        sink_broadcast_size[axis] = module.weight.size(axis)
         if isinstance(module, _batch_norm):
             # We re-compute the bias as function of running_mean and running_var to adjust the
             # additive factor for equalization.
@@ -466,7 +478,7 @@ def _no_equalize():
                 module, module.bias.clone() + additive_factor * (scaling_factors - 1), attr='bias')
         _update_weights(
             module,
-            module.weight.clone() * torch.reshape(scaling_factors, src_broadcast_size),
+            module.weight.clone() * torch.reshape(scaling_factors, sink_broadcast_size),
             attr='weight')
 
     return scaling_factors
@@ -575,6 +587,8 @@ def find_srcs(graph_model: GraphModule, starting_node: Node,
               node.op == 'call_function' and node.target in _residual_fns):
             find_srcs(graph_model, node, state)
             find_sinks(graph_model, node, state)
+        elif node.target in _ignore_ops:
+            continue
         else:
             # If we meet an unrecognized op, we add None to invalidate the region
             state.srcs.add(_UNSUPPORTED_OP)
@@ -606,6 +620,8 @@ def find_sinks(graph_model: GraphModule, starting_node: Node,
               node.op == 'call_function' and node.target in _residual_fns):
             find_sinks(graph_model, node, state)
             find_srcs(graph_model, node, state)
+        elif node.target in _ignore_ops:
+            continue
         else:
             # If we meet an unrecognized op, we add None to invalidate the region
             state.sinks.add(_UNSUPPORTED_OP)
@@ -713,7 +729,7 @@ def setup(self):
         for region in self.regions:
             batch_dim = 0
             if hasattr(region, 'batch_first'):
-                batch_dim = 0 if region.batch_first == True else 1
+                batch_dim = 0 if region.batch_first else 1
 
             hook_fn = partial(
                 self.forward_stats_hook, name=region, batch_dim=batch_dim, use_inp=True)
@@ -761,16 +777,14 @@ def forward_stats_hook(self, module, *args, name, batch_dim=0, use_inp=True, **k
         # Extra check for batch_dim
         if hasattr(x, 'names') and 'N' in x.names:
             batch_dim = x.names.index('N')
-            x = x.transpose(0, batch_dim)
 
         self.batch_dim_act_map[name] = batch_dim
 
+        input_scales = self.scale_fn(x, dim=batch_dim)
         if name not in self.float_act_map:
-            self.float_act_map[name] = self.scale_fn(x, dim=batch_dim)
+            self.float_act_map[name] = input_scales
         else:
-            batch_data = torch.cat([self.float_act_map[name].unsqueeze(batch_dim), x],
-                                   dim=batch_dim)
-            self.float_act_map[name] = self.scale_fn(batch_data, dim=batch_dim)
+            self.float_act_map[name] = torch.max(self.float_act_map[name], input_scales)
 
     def insert_mul_node(self, scale, shape, axis, region, batch_dim=0):
         broadcastable_shape = [1] * len(shape)
@@ -832,7 +846,7 @@ def setup(self):
                 for name in region.srcs + region.sinks:
                     module = name_to_module[name]
                     if hasattr(module, 'batch_first'):
-                        batch_dim = 0 if module.batch_first == True else 1
+                        batch_dim = 0 if module.batch_first else 1
                 for name in region_to_search:
                     act_module = name_to_module[name]
                     use_inp = True if region_to_search == region.sinks else False
@@ -907,16 +921,12 @@ def forward_stats_hook(self, module, *args, name, batch_dim=0, use_inp=True, **k
         # Extra check for batch_dim
         if hasattr(x, 'names') and 'N' in x.names:
             batch_dim = x.names.index('N')
-            x = x.transpose(0, batch_dim)
-
-        self.batch_dim_act_map[name] = batch_dim
 
+        input_scales = self.scale_fn(x, dim=batch_dim)
         if name not in self.float_act_map:
-            self.float_act_map[name] = self.scale_fn(x, dim=batch_dim)
+            self.float_act_map[name] = input_scales
         else:
-            batch_data = torch.cat([self.float_act_map[name].unsqueeze(batch_dim), x],
-                                   dim=batch_dim)
-            self.float_act_map[name] = self.scale_fn(batch_data, dim=batch_dim)
+            self.float_act_map[name] = torch.max(self.float_act_map[name], input_scales)
 
     def insert_mul_node(self, scale, shape, axis, act_node, batch_dim=0):
         broadcastable_shape = [1] * len(shape)

src/brevitas/graph/standardize.py

@@ -107,7 +107,8 @@ class TorchFunctionalToModule(GraphTransform):
                                                        nn.AvgPool1d), (F.avg_pool2d, nn.AvgPool2d),
                         (F.avg_pool3d, nn.AvgPool3d), (F.adaptive_avg_pool1d, nn.AdaptiveAvgPool1d),
                         (F.adaptive_avg_pool2d,
-                         nn.AdaptiveAvgPool2d), (F.adaptive_avg_pool3d, nn.AdaptiveAvgPool3d))
+                         nn.AdaptiveAvgPool2d), (F.adaptive_avg_pool3d,
+                                                 nn.AdaptiveAvgPool3d), (F.dropout, nn.Dropout))
 
     def __init__(self, fn_to_module_map=FN_TO_MODULE_MAP):
         super().__init__()

src/brevitas_examples/llm/llm_quant/equalize.py

@@ -10,6 +10,8 @@
 from brevitas.fx.brevitas_tracer import value_trace
 from brevitas.graph.equalize import activation_equalization_mode
 from brevitas.graph.equalize import EqualizeGraph
+from brevitas.graph.standardize import DuplicateSharedStatelessModule
+from brevitas.graph.standardize import TorchFunctionalToModule
 from brevitas_examples.llm.llm_quant.run_utils import apply_layer_ptq_fn
 from brevitas_examples.llm.llm_quant.run_utils import cast_to_float32
 
@@ -26,6 +28,13 @@ def activation_equalization_iter(curr_layer, inps, outs, cached_values, alpha):
     return outs
 
 
+def trace_and_standardize(model, ref_kwargs):
+    graph_model = value_trace(model, value_args=ref_kwargs)
+    graph_model = TorchFunctionalToModule().apply(graph_model)
+    graph_model = DuplicateSharedStatelessModule().apply(graph_model)
+    return graph_model
+
+
 @torch.no_grad()
 def apply_act_equalization(
         model,
@@ -49,7 +58,7 @@ def apply_act_equalization(
         # We can't do fp16 tracing on CPU as many kernels are not implemented
         # So we have to cast to fp32 first, trace, apply equalization, and then cast back
         with cast_to_float32(model, dtype):
-            graph_model = value_trace(model, value_args=ref_kwargs)
+            graph_model = trace_and_standardize(model, ref_kwargs=ref_kwargs)
             # TODO this is currently running on CPU. We need Accelerate or a TorchDispatchMode
             # or an FX interpreter to run it on GPU
             warnings.warn(
@@ -70,5 +79,5 @@ def apply_weight_equalization(model, dtype, ref_kwargs, scale_computation_type='
     # We can't do fp16 tracing on CPU as many kernels are not implemented
     # So we have to cast to fp32 first, trace, apply equalization, and then cast back
     with cast_to_float32(model, dtype):
-        graph_model = value_trace(model, value_args=ref_kwargs)
+        graph_model = trace_and_standardize(model, ref_kwargs=ref_kwargs)
         EqualizeGraph(scale_computation_type=scale_computation_type).apply(graph_model)

src/brevitas_examples/llm/main.py

@@ -303,9 +303,14 @@ def main():
             quantize_embedding=args.quantize_embedding,
             seqlen=args.seqlen)
         # Tie back first/last layer weights in case they got untied
-        model.tie_weights()
         print("Model quantization applied.")
 
+    # If any equalization has taken places, the embedding layer and the fully connected one are
+    # not tied anymore, and they need to be treated as standalone, separate layers.
+    # In all other cases we can tie them back so to preserve memory.
+    if args.act_equalization is None and not args.weight_equalization:
+        model.tie_weights()
+
     if args.act_calibration:
         print("Apply act calibration...")
         apply_calibration(model, calibration_loader, args.nsamples)