AWQ Apply Scales Bugfix when smooth layer output length doesn't match balance layer input length (#1451)

### Summary
We are hitting an edge case in AWQ we had not previously hit with the
initial Llama/Qwen testing models. When a smooth layer's # of
output_features does not match a balance layer's # of input_features,
the code as it is currently will error out when trying to update the
smooth layer's weights with `weights.div(scales)`, due to a shape
mismatch error. We are hitting this in #1440 for Phi3 models, which
include a mapping between the fused `qkv_proj` smooth layer and `o_proj`
balance layer in AutoAWQ (see
[here](https://github.com/casper-hansen/AutoAWQ/blob/main/awq/models/phi3.py#L51-L57)).

The resolution in AutoAWQ is to only use the last rows of the smooth
layer so that the shapes line up, as shown
[here](https://github.com/casper-hansen/AutoAWQ/blob/main/awq/quantize/scale.py#L123).
This PR includes that update, and with #1440 will allow Phi3 models to
be quantizable with AWQModifier. Like with v_proj -> o_proj, if shapes
don't match up, they will be excluded from resolved mappings. This
allows
[phi-3-mini](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/tree/main?show_file_info=model-00001-of-00002.safetensors)
to include the mapping because `qkv_proj out_features == 3*o_proj
in_features == 9216`, but excludes it from
[phi-3-medium](https://huggingface.co/microsoft/Phi-3-medium-128k-instruct/tree/main?show_file_info=model-00001-of-00006.safetensors)
which has `qkv_proj out_features == 7680` and `o_proj
in_features==5120`. If the mapping is included for phi-3-medium, the
model blows up with wikitext eval perplexities >2000. This
implementation was agreed upon with @anmarques .

PS: I also moved `mul` & `div` to `mul_` & `div_`, to avoid unnecessary
memory allocation.

-------------
### Test Plan
With these changes and with #1440 , `examples/awq/llama_example.py`
works with `"microsoft/Phi-3-mini-128k-instruct"` and produces similar
results as when qkv_proj to o_proj mapping is included

Without mapping:

| Tasks  |Version|Filter|n-shot|    Metric     |   | Value |   |Stderr|
|--------|------:|------|-----:|---------------|---|------:|---|------|
|wikitext|      2|none  |     5|bits_per_byte  |↓  | 0.6474|±  |   N/A|
|        |       |none  |     5|byte_perplexity|↓  | 1.5664|±  |   N/A|
|        |       |none  |     5|word_perplexity|↓  |11.0201|±  |   N/A|

With mapping:

| Tasks  |Version|Filter|n-shot|    Metric     |   | Value |   |Stderr|
|--------|------:|------|-----:|---------------|---|------:|---|------|
|wikitext|      2|none  |     5|bits_per_byte  |↓  | 0.6482|±  |   N/A|
|        |       |none  |     5|byte_perplexity|↓  | 1.5672|±  |   N/A|
|        |       |none  |     5|word_perplexity|↓  |11.0527|±  |   N/A|

I also confirmed re-running with `meta-llama/Llama-3.2-3B-Instruct` and
`meta-llama/Llama-2-7b-hf` does not deviate in PPL scores from what is
currently on `main`

---------

Signed-off-by: Brian Dellabetta <bdellabe@redhat.com>

Author: brian-dellabetta

src/llmcompressor/modifiers/awq/base.py

@@ -310,14 +310,24 @@ def _set_resolved_mappings(self, model: Module) -> None:
                         if not balance_layer:
                             continue
 
-                        # exclude v_proj/o_proj mappings whose shapes are incompatible
+                        # exclude v_proj->o_proj mappings whose shapes are incompatible
                         # https://github.com/mit-han-lab/llm-awq/pull/67#issuecomment-1681632777
                         if (
-                            ".v_proj" in layer_name
-                            and ".o_proj" in balance_name
-                            and isinstance(smooth_layer, torch.nn.Linear)
+                            isinstance(smooth_layer, torch.nn.Linear)
                             and isinstance(balance_layer, torch.nn.Linear)
-                            and smooth_layer.weight.shape != balance_layer.weight.shape
+                            and ".o_proj" in balance_name
+                            and (
+                                (
+                                    ".v_proj" in layer_name
+                                    and smooth_layer.out_features
+                                    != balance_layer.in_features
+                                )
+                                or (
+                                    ".qkv_proj" in layer_name
+                                    and smooth_layer.out_features
+                                    != 3 * balance_layer.in_features
+                                )
+                            )
                         ):
                             num_skipped_oproj_mappings += 1
                             continue
@@ -466,33 +476,42 @@ def _apply_smoothing(self, model: Module) -> None:
                     inp, w_mean, x_mean, module2inspect, balance_layers, fp16_output
                 )
 
-            scales = best_scales
-
             @torch.no_grad()
             def smooth(module):
                 with align_module_device(module):
+                    scales = best_scales.to(module.weight.device)
                     if module in balance_layers:
-                        module.weight.mul_(scales.view(1, -1).to(module.weight.device))
+                        update_offload_parameter(
+                            module,
+                            "weight",
+                            module.weight.mul_(scales.view(1, -1)),
+                        )
                     elif module == smooth_layer:
                         if module.weight.ndim == 1:
                             update_offload_parameter(
                                 module,
                                 "weight",
-                                module.weight.div(scales.to(module.weight.device)),
+                                module.weight.div_(scales),
                             )
                         else:
+                            # NOTE: edge case when smooth layer number of out_features
+                            # is not equal to balance layer number of in_features
+                            # e.g. when fused qkv_proj is used to smooth o_proj
+                            # in this case, default to scaling the last output features
+                            # because the desired smooth layer is v_proj
+                            # https://github.com/casper-hansen/AutoAWQ/blob/main/awq/quantize/scale.py#L123
                             update_offload_parameter(
                                 module,
                                 "weight",
-                                module.weight.div(
-                                    scales.view(-1, 1).to(module.weight.device)
+                                module.weight[-scales.size(0) :].div_(
+                                    scales.view(-1, 1)
                                 ),
                             )
                         if hasattr(module, "bias") and module.bias is not None:
                             update_offload_parameter(
                                 module,
                                 "bias",
-                                module.bias.div(scales.to(module.bias.device)),
+                                module.bias.div_(scales),
                             )
 
             parent = get_fsdp_parent(mapping.smooth_name, model)