Mixtral scaling: Reduce perplexity from 4.294 to 4.269

awq/models/base.py

@@ -35,6 +35,7 @@
 
 from awq.models._config import AwqConfig
 from awq.modules.act import ScaledActivation
+from awq.modules.moe import ScaledMixtralSparseMoeBlock
 from awq.modules.linear import WQLinear_GEMM, WQLinear_GEMV
 from awq.modules.exllama import WQLinear_Exllama
 from awq.modules.exllamav2 import WQLinear_ExllamaV2
@@ -96,11 +97,13 @@ def quantize(
         split="train",
         text_column="text",
         duo_scaling=True,
-        modules_to_not_convert=None,
         export_compatible=False,
     ):
         self.quant_config: AwqConfig = AwqConfig.from_dict(quant_config)
 
+        if hasattr(self, "modules_to_not_convert"):
+            self.quant_config.modules_to_not_convert = self.modules_to_not_convert
+
         self.quantizer = AwqQuantizer(
             self,
             self.model,
@@ -112,7 +115,7 @@ def quantize(
             split,
             text_column,
             duo_scaling,
-            modules_to_not_convert=modules_to_not_convert,
+            modules_to_not_convert=self.quant_config.modules_to_not_convert,
             export_compatible=export_compatible,
         )
         self.quantizer.quantize()
@@ -402,6 +405,9 @@ def _load_quantized_modules(
             # Replace activation functions
             self._scale_activations(self, layer)
 
+            # Replace mixture of experts
+            self._scale_moe(self, layer)
+
             # Replace nn.Linear with WQLinear
             for name, module in named_linears.items():
                 if use_exllama:
@@ -436,5 +442,19 @@ def _scale_activations(self, layer):
                 )
 
                 # scale activation
-                scaled_act = ScaledActivation(scale_dict["scale_layer"], scale_like)
-                set_op_by_name(layer, scale_dict["scale_name"], scaled_act)
+                scaled_act = ScaledActivation(scale_dict['scale_layer'], scale_like)
+                set_op_by_name(layer, scale_dict['scale_name'], scaled_act)
+
+    def _scale_moe(self, layer):
+        if hasattr(self, "get_moe_for_scaling") and hasattr(layer.block_sparse_moe, "scales"):
+            scale_dict: dict = self.get_moe_for_scaling(layer)
+
+            if not isinstance(scale_dict['scale_layer'], ScaledMixtralSparseMoeBlock):
+                param = next(layer.parameters())
+
+                # get activation scale
+                scale_like = torch.ones(scale_dict['scale_shape'], dtype=param.dtype, device=param.device)
+
+                # scale moe
+                scaled_act = ScaledMixtralSparseMoeBlock(scale_dict['scale_layer'], scale_like)
+                set_op_by_name(layer, scale_dict['scale_name'], scaled_act)

awq/models/mixtral.py

@@ -6,13 +6,26 @@
 from awq.modules.fused.model import MixtralModel
 from transformers.models.mixtral.modeling_mixtral import (
     MixtralDecoderLayer as OldMixtralDecoderLayer,
-    MixtralForCausalLM as OldMixtralForCausalLM
+    MixtralForCausalLM as OldMixtralForCausalLM,
+    MixtralBLockSparseTop2MLP as OldMixtralBLockSparseTop2MLP,
 )
 from awq.modules.fused.norm import FasterTransformerRMSNorm
 
+def _transformers_version_check():
+    import transformers
+    tv = transformers.__version__.split('.')
+    if len(tv) == 4:
+        major, minor, patch, dev = tv
+    else:
+        major, minor, patch = tv
+
+    if int(major) == 4 and int(minor) < 37:
+        raise Exception("Mixtral requires a minimum of 4.37.0.dev0: pip install git+https://github.com/huggingface/transformers.git")
+
 class MixtralAWQForCausalLM(BaseAWQForCausalLM):
     layer_type = "MixtralDecoderLayer"
     max_new_tokens_key = "max_position_embeddings"
+    modules_to_not_convert = ["gate"]
 
     @staticmethod
     def fuse_layers(model: OldMixtralForCausalLM):
@@ -21,6 +34,7 @@ def fuse_layers(model: OldMixtralForCausalLM):
 
     @staticmethod
     def get_model_layers(model: OldMixtralForCausalLM):
+        _transformers_version_check()
         return model.model.layers
 
     @staticmethod
@@ -29,6 +43,14 @@ def get_act_for_scaling(module):
             is_scalable=False
         )
 
+    @staticmethod
+    def get_moe_for_scaling(module: OldMixtralDecoderLayer):
+        return dict(
+            scale_name="block_sparse_moe",
+            scale_layer=module.block_sparse_moe,
+            scale_shape=(module.block_sparse_moe.num_experts, module.block_sparse_moe.hidden_dim),
+        )
+
     @staticmethod
     def move_embed(model: OldMixtralForCausalLM, device: str):
         model.model.embed_tokens = model.model.embed_tokens.to(device)
@@ -53,19 +75,18 @@ def get_layers_for_scaling(module: OldMixtralDecoderLayer, input_feat, module_kw
                 layers=[module.self_attn.o_proj],
                 inp=input_feat['self_attn.o_proj'],
             ))
-
-        # linear in
+
+        # NOTE: Scaled in awq.quantize.scale.scale_moe_experts, awq.modules.moe.ScaledMixtralSparseMoeBlock
+        # Experts: Not a linear layer, special handling is introduced in awq.quantize.quantizer
         layers.append(dict(
-            prev_op=module.post_attention_layernorm,
-            layers=[
-                w for expert in module.block_sparse_moe.experts
-                  for w in [expert.w1, expert.w3]
-            ],
+            prev_op=module.block_sparse_moe,
+            layers=module.block_sparse_moe.experts,
             inp=input_feat['block_sparse_moe'],
             module2inspect=module.block_sparse_moe,
         ))
 
-        # linear out
+        # scaling w2        
+        expert: OldMixtralBLockSparseTop2MLP
         for i, expert in enumerate(module.block_sparse_moe.experts):
             layers.append(dict(
                 prev_op=expert.w3,

awq/modules/moe.py

@@ -0,0 +1,75 @@
+import torch
+from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
+
+class ScaledMixtralSparseMoeBlock(torch.nn.Module):
+    """
+    This is a modified sparse MoE that scales experts individually.
+
+    Modified version of:
+    transformers.models.mixtral.modeling_mixtral.MixtralSparseMoeBlock
+    """
+
+    def __init__(self, prev_op: MixtralSparseMoeBlock, scales: torch.Tensor):
+        super().__init__()
+        self.hidden_dim = prev_op.hidden_dim
+        self.ffn_dim = prev_op.ffn_dim
+        self.num_experts = prev_op.num_experts
+        self.top_k = prev_op.top_k
+
+        # gating
+        self.gate = prev_op.gate
+
+        # experts
+        self.experts = prev_op.experts
+
+        # [expert_num, hidden_dim]
+        self.scales = torch.nn.Parameter(scales.data)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+
+        routing_weights = torch.nn.functional.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        # we cast back to the input dtype
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+
+            ### NOTE: We scale weights here, modified from original MoE.
+            current_state = hidden_states[None, top_x_list].reshape(
+                -1, hidden_dim) / self.scales[expert_idx] # <-- scales
+
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states, router_logits

awq/quantize/quantizer.py

@@ -4,9 +4,10 @@
 import functools
 import torch.nn as nn
 from tqdm import tqdm
-from typing import Dict, List
 from collections import defaultdict
+from typing import Dict, List, Union
 from awq.utils.utils import clear_memory
+from transformers import PreTrainedModel
 from awq.utils.calib_data import get_calib_dataset
 from awq.quantize.scale import apply_scale, apply_clip
 from awq.modules.linear import WQLinear_GEMM, WQLinear_GEMV
@@ -17,14 +18,15 @@
     set_op_by_name,
     exclude_layers_to_not_quantize
 )
+from transformers.models.mixtral.modeling_mixtral import MixtralBLockSparseTop2MLP
 
 
 class AwqQuantizer:
     def __init__(self, awq_model, model, tokenizer, w_bit, group_size, version, 
                        calib_data, split, text_column, duo_scaling, modules_to_not_convert=None,
                        export_compatible=False) -> None:
         self.awq_model = awq_model
-        self.model = model
+        self.model: PreTrainedModel = model
         self.tokenizer = tokenizer
         self.w_bit = w_bit
         self.group_size = group_size
@@ -162,7 +164,8 @@ def _apply_quant(self, module, named_linears: Dict[str, nn.Linear]):
             clear_memory()
 
     @torch.no_grad()
-    def _search_best_scale(self, module, prev_op, layers: List[nn.Linear], inp: torch.Tensor, module2inspect=None, kwargs={}):
+    def _search_best_scale(self, module, prev_op, layers: Union[List[nn.Linear], List[MixtralBLockSparseTop2MLP]],
+                           inp: torch.Tensor, module2inspect=None, kwargs={}):
         if module2inspect is None:
             assert len(layers) == 1
             module2inspect = layers[0]
@@ -174,13 +177,23 @@ def _search_best_scale(self, module, prev_op, layers: List[nn.Linear], inp: torc
         inp = inp.to(next(module2inspect.parameters()).device)
 
         # [STEP 1]: Compute maximum of weight
-        weight = torch.cat([_m.weight for _m in layers], dim=0)
-        org_shape = weight.shape
-        weight = weight.view(-1, self.group_size)
-        w_scale = weight.abs() / weight.abs().amax(dim=1, keepdim=True)
-        w_scale = w_scale.view(org_shape)
-        w_max = w_scale.mean(0)
-        clear_memory(weight)
+        def _get_w_max(layer_weights):
+            weight = torch.cat([_m.weight for _m in layer_weights], dim=0)
+            org_shape = weight.shape
+            weight = weight.view(-1, self.group_size)
+            w_scale = weight.abs() / weight.abs().amax(dim=1, keepdim=True)
+            w_scale = w_scale.view(org_shape)
+            w_max = w_scale.mean(0)
+            clear_memory(weight)
+
+            return w_max
+
+        if type(layers[0]) == nn.Linear:
+            w_max = _get_w_max(layers)
+        else:
+            # FIXME: Specific to Mixtral
+            weights = [[expert.w1, expert.w3] for expert in layers]
+            w_max = [_get_w_max(weight) for weight in weights]
 
         # [STEP 2]: Compute maximum of x
         x_max = inp.abs().view(-1, inp.shape[-1]).mean(0)
@@ -194,12 +207,23 @@ def _search_best_scale(self, module, prev_op, layers: List[nn.Linear], inp: torc
                 fp16_output = fp16_output[0]
 
         # [STEP 4]: Compute loss
-        best_scales = self._compute_best_scale(
-            inp, w_max, x_max, module2inspect,
-            layers, fp16_output, module_kwargs
-        )
+        if type(layers[0]) == nn.Linear:
+            best_scales = self._compute_best_scale(
+                inp, w_max, x_max, module2inspect,
+                layers, fp16_output, module_kwargs
+            )
+        else:
+            best_scales = [
+                self._compute_best_scale(
+                    inp, w_max[i], x_max, module2inspect,
+                    experts, fp16_output, module_kwargs
+                ) for i, experts in enumerate(weights)
+            ]
+
+        prev_op_name = get_op_name(module, prev_op)
+        layer_names = tuple([get_op_name(module, m) for m in layers])
 
-        return (get_op_name(module, prev_op), tuple([get_op_name(module, m) for m in layers]), best_scales)
+        return (prev_op_name, layer_names, best_scales)
 
     def _compute_best_scale(self, x, w_max, x_max, module2inspect, linears2scale: List[nn.Linear],
                                   fp16_output, kwargs={}):

awq/quantize/scale.py

@@ -2,13 +2,19 @@
 import torch.nn as nn
 from typing import Tuple, List
 from awq.modules.act import ScaledActivation
+from awq.modules.moe import ScaledMixtralSparseMoeBlock
 from awq.utils.module import get_op_by_name, set_op_by_name
 from transformers.models.bloom.modeling_bloom import BloomGelu
 from transformers.models.llama.modeling_llama import LlamaRMSNorm
 from transformers.activations import NewGELUActivation, PytorchGELUTanh, GELUActivation
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralSparseMoeBlock,
+    MixtralBLockSparseTop2MLP,
+)
 
 allowed_norms = [nn.LayerNorm, LlamaRMSNorm]
 allowed_act_fns = [nn.GELU, BloomGelu, NewGELUActivation, PytorchGELUTanh, GELUActivation]
+allowed_moe = [MixtralSparseMoeBlock]
 
 @torch.no_grad()
 def apply_clip(module, clip_list: Tuple[str, torch.Tensor]):
@@ -31,7 +37,12 @@ def apply_scale(module, scales_list, input_feat_dict=None):
         prev_op.cuda()
         for layer in layers:
             layer.cuda()
-        scales.cuda()
+
+        if type(scales) == list:
+            for scale in scales:
+                scale.cuda()
+        else:
+            scales.cuda()
 
         if isinstance(prev_op, nn.Linear) and type(layers) == list and isinstance(layers[0], nn.Linear):
             scale_fc_fcs(prev_op, layers, scales)
@@ -48,6 +59,15 @@ def apply_scale(module, scales_list, input_feat_dict=None):
             new_module = ScaledActivation(prev_op, scales)
             set_op_by_name(module, prev_op_name, new_module)
             scale_gelu_fc(prev_op, layers[0], scales)
+
+        elif any(isinstance(prev_op,t) for t in allowed_moe):
+            # scales: [best_scale_expert_0, best_scale_expert_1, ...] -> [expert_index, scales]
+            scales = torch.stack(scales).cuda()
+
+            # apply scales
+            new_module = ScaledMixtralSparseMoeBlock(prev_op, scales)
+            set_op_by_name(module, prev_op_name, new_module)
+            scale_moe_experts(prev_op, layers, scales)
 
         else:
             raise NotImplementedError(
@@ -64,7 +84,12 @@ def apply_scale(module, scales_list, input_feat_dict=None):
         prev_op.cpu()
         for layer in layers:
             layer.cpu()
-        scales.cpu()
+
+        if type(scales) == list:
+            for scale in scales:
+                scale.cpu()
+        else:
+            scales.cpu()
 
 @torch.no_grad()
 def scale_ln_fcs(ln: nn.Linear, fcs: List[nn.Linear], scales: torch.Tensor):
@@ -133,3 +158,18 @@ def scale_gelu_fc(gelu: allowed_act_fns, fc: nn.Linear, scales: torch.Tensor):
 
     for p in fc.parameters():
         assert torch.isnan(p).sum() == 0
+
+@torch.no_grad()
+def scale_moe_experts(moe: MixtralSparseMoeBlock, experts: List[MixtralBLockSparseTop2MLP], scales: torch.Tensor):
+    assert any(isinstance(moe, allowed_module) for allowed_module in allowed_moe)
+    assert all(isinstance(m, MixtralBLockSparseTop2MLP) for m in experts)
+
+    # One scale for each expert, applied to w1 and w3 only
+    # Not applied to w2 because it does not take hidden_states as input
+    for i, expert in enumerate(experts):
+        expert.w1.weight.mul_(scales[i].view(1, -1))
+        expert.w3.weight.mul_(scales[i].view(1, -1))
+
+    for expert in experts:
+        for p in expert.parameters():
+            assert torch.isnan(p).sum() == 0