Add SpinQuant to generate.py (#1069)

* Only import SpinQuant when necessary

No need to import the large Hadamard matrices required for SpinQuant if it isn't necessary

* Add SpinQaunt to `generate.py`

* Custom op for Hadamard transform for torch.compile compatability

* Add spinquant to arg parser info

* Add Spinquant benchmark results to README

* Add performance testing details

* Fix broken custom op for PyTorch < 2.4

Author: tobiasvanderwerff

torchao/_models/llama/eval.py

@@ -31,7 +31,6 @@
 from tokenizer import get_tokenizer
 import time
 from torchao.utils import TORCH_VERSION_AT_LEAST_2_5
-from torchao.prototype.spinquant import apply_spinquant
 
 def run_evaluation(
     checkpoint_path: Path,
@@ -71,6 +70,7 @@ def run_evaluation(
 
     if quantization:
         if "spinquant" in quantization:
+            from torchao.prototype.spinquant import apply_spinquant
             apply_spinquant(model)
         if "int8wo" in quantization:
             quantize_(model, int8_weight_only())

torchao/_models/llama/generate.py

@@ -217,6 +217,9 @@ def main(
             float8_dynamic_activation_float8_weight,
         )
         from torchao.quantization.granularity import PerTensor, PerRow
+        if "spinquant" in quantization:
+            from torchao.prototype.spinquant import apply_spinquant
+            apply_spinquant(model)
         if "int8wo" in quantization:
             quantize_(model, int8_weight_only())
         if "int8dq" in quantization:
@@ -460,7 +463,7 @@ def callback(x):
     parser.add_argument('-q', '--quantization', type=str, 
         help=(
             'Which quantization techniques to apply: int8dq, int8wo, fp6, int4wo-<groupsize>, int4wo-<groupsize>-hqq, autoquant, '
-            +'autoquant-int4, autoquant-float8, uintx-<nbits>-<groupsize>, uintx-<nbits>-<groupsize>-hqq, sparse-marlin'
+            +'autoquant-int4, autoquant-float8, uintx-<nbits>-<groupsize>, uintx-<nbits>-<groupsize>-hqq, sparse-marlin, spinquant'
         )
     )
     parser.add_argument("--calibration_limit", type=int, default=10, help="Number of calibration examples")

torchao/prototype/spinquant/README.md

@@ -4,8 +4,37 @@ Re-implementation of SpinQuant based on the official code implementation (https:
 
 ## Usage
 
-Using this implementation with CUDA requires installing the Fast Hadamard Transform CUDA package, which can be done as follows:
+For optimal performance on CUDA GPUs, install the Fast Hadamard Transform package:
 
 ```shell
 pip install git+https://github.com/Dao-AILab/fast-hadamard-transform.git
-```
+```
+
+## Performance 
+
+See https://github.com/pytorch/ao/pull/983 for Wikitext benchmark results.
+
+Tested on:
+
+- Llama-2-7b
+- PyTorch 2.4.1
+- NVIDIA A100
+- CUDA 12.1
+
+Without `torch.compile`:
+
+| Configuration  | Average tokens/sec | Average Bandwidth (GB/s) | Peak Memory Usage (GB) | Model Size (GB) |
+|----------------|--------------------|--------------------------|------------------------|-----------------|
+| Baseline       | 27.33              | 361.21                   | 13.62                  | 13.21           |
+| Spinquant (R4) | 23.01              | 304.10                   | 14.24                  | 13.22           |
+
+With `torch.compile`:
+
+| Configuration        | Average tokens/sec | Average Bandwidth (GB/s) | Peak Memory Usage (GB) | Model Size (GB) |
+|----------------------|--------------------|--------------------------|------------------------|-----------------|
+| Baseline             | 114.08             | 1507.58                  | 13.88                  | 13.21           |
+| Spinquant (R4)       | 109.59             | 1448.61                  | 13.72                  | 13.22           |
+| Spinquant (R1+R2+R4) | 109.64             | 1449.28                  | 14.90                  | 13.22           |
+
+
+NB: R1 and R2 are fused into the linear weights before inference takes place, so it is expected that they do not lead to additional overhead at inference time.

torchao/prototype/spinquant/hadamard_utils.py

@@ -11,10 +11,12 @@
 
 import torch
 
+from torchao.ops import lib
 from torchao.prototype.spinquant._hadamard_matrices import get_had172, get_had156, get_had140, get_had108, get_had60, get_had52, get_had36, get_had28, get_had44, get_had40, get_had20, get_had12
+from torchao.utils import TORCH_VERSION_AT_LEAST_2_4
 
 try:
-    from fast_hadamard_transform import hadamard_transform
+    from fast_hadamard_transform import hadamard_transform as _fast_hadamard_transform
 
     def matmul_hadU(X, hadK, K):
         if X.is_cuda:
@@ -32,16 +34,59 @@ def matmul_hadU(X, hadK, K):
         return matmul_hadU_slow(X, hadK, K)
 
 
+def register_custom_op_impl(name):
+    def decorator(func):
+        if TORCH_VERSION_AT_LEAST_2_4:
+            return torch.library.custom_op(f"{name}", mutates_args=())(func)
+        else:
+            lib.define("hadamard_transform(Tensor x, float scale = 0.0) -> Tensor")
+            return torch.library.impl(f"{name}", "cuda")(func)
+    return decorator
+
+
+def register_custom_op_abstract(name):
+    def decorator(func):
+        if TORCH_VERSION_AT_LEAST_2_4:
+            return torch.library.register_fake(f"{name}")(func)
+        else:
+            return torch.library.impl_abstract(f"{name}")(func)
+    return decorator
+
+
+@register_custom_op_impl("torchao::hadamard_transform")
+def hadamard_transform(x: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+    """
+    Arguments:
+        x: (..., dim)
+        scale: float. Multiply the output by this number.
+    Returns:
+        out: (..., dim)
+
+    Multiply each row of x by the Hadamard transform matrix.
+    Equivalent to F.linear(x, torch.tensor(scipy.linalg.hadamard(dim))) * scale.
+    If dim is not a power of 2, we implicitly pad x with zero so that dim is the next power of 2.
+
+    Source: https://github.com/Dao-AILab/fast-hadamard-transform
+    """
+    return _fast_hadamard_transform(x, scale)
+
+
+@register_custom_op_abstract("torchao::hadamard_transform")
+def _(x: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+    torch._check(x.dim() >= 1, lambda: f"input should be at least a 1D tensor, got {x.dim()}D")
+    return torch.empty_like(x)
+
+
 class HadamardTransform(torch.autograd.Function):
     """The unnormalized Hadamard transform (i.e. without dividing by sqrt(2))"""
 
     @staticmethod
     def forward(ctx, u):
-        return hadamard_transform(u)
+        return _fast_hadamard_transform(u)
 
     @staticmethod
     def backward(ctx, grad):
-        return hadamard_transform(grad)
+        return _fast_hadamard_transform(grad)
 
 
 def is_pow2(n):
@@ -144,9 +189,9 @@ def matmul_hadU_slow(X, hadK, K):
 def matmul_hadU_fast(X, hadK, K):
     n = X.shape[-1]
     if K == 1:
-        return HadamardTransform.apply(X.contiguous()) / torch.tensor(n).sqrt()
+        return torch.ops.torchao.hadamard_transform.default(X.contiguous()) / torch.tensor(n).sqrt()
     input = X.view(-1, K, n // K)
-    input = HadamardTransform.apply(input.contiguous()) / torch.tensor(n).sqrt()
+    input = torch.ops.torchao.hadamard_transform.default(input.contiguous()) / torch.tensor(n).sqrt()
     input = hadK.to(input.device).to(input.dtype) @ input
     return input.reshape(X.shape)
 

torchao/prototype/spinquant/spinquant.py

@@ -103,7 +103,7 @@ def apply_spinquant_r4(model, device):
     _add_activation_wrappers_r4(model)
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def _fuse_layernorm_into_linear(layernorm: RMSNorm, linear_layers: typing.Iterable[torch.nn.Linear]):
     """Fuse the linear operations in Layernorm into the adjacent linear blocks."""
     for linear in linear_layers:
@@ -127,7 +127,7 @@ def _fuse_layernorm_into_linear(layernorm: RMSNorm, linear_layers: typing.Iterab
     layernorm.weight.data = torch.ones_like(layernorm.weight.data)
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def _rotate_model_r1(model, R1):
     _rotate_embeddings(model, R1)
     _rotate_head(model, R1)
@@ -139,7 +139,7 @@ def _rotate_model_r1(model, R1):
         _rotate_mlp_output(layer, R1)
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def _rotate_model_r2(model, R2s):
     """Rotate the W_v and W_o weights of the multi-head self-attention modules."""
 
@@ -168,7 +168,7 @@ def _rotate_model_r2(model, R2s):
         attn.wqkv.weight.data = torch.cat([wq, wk, wv_mod.weight.data], dim=0)
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def _rotate_model_r4(model):
     """Rotate the MLP output weights."""
 
@@ -193,7 +193,7 @@ def _add_activation_wrappers_r4(model):
         )
 
 
-@torch.inference_mode()
+@torch.no_grad()
 def fuse_layernorm_into_linear(model):
     """
     Fuse RMSNorm weights into the subsequent linear layers.