Merge pull request #62 from andrea-fasoli/unit_test_int8

test: Unit test int8

Author: chichun-charlie-liu

fms_mo/aiu_addons/gptq/gptq_aiu_linear.py

@@ -141,7 +141,7 @@ def get_gptq_aiu_linear(
     in_features: int,
     out_features: int,
     bias: bool,
-    linear_config: Optional[Mapping[str, Any]] = None,
+    linear_config: Mapping[str, Any],
 ) -> torch.nn.Module:
     """Retrieve a GPTQ W4A16 Linear module"""
 

fms_mo/aiu_addons/i8i8/i8i8_aiu_adapter.py

@@ -47,8 +47,8 @@ def _int8_qparams_aiu(
 
 
 def _add_defaults_and_concat(
-    new_sd: Mapping[str, torch.Tensor],
-    modules_seen: set,
+    new_sd: dict[str, torch.Tensor],
+    modules_seen: set[str],
 ) -> None:
     """
     Add default activation clip values, zero_shift, and smoothquant_scale (if not

fms_mo/aiu_addons/i8i8/i8i8_aiu_op.py

@@ -114,7 +114,7 @@ def extract_qdata(
     w_in_feat: int,
     w_out_feat: int,
     smoothquant: bool,
-) -> tuple[torch.Tensor]:
+) -> tuple[torch.Tensor, ...]:
     """6 tensors are to be de-concatenated from qdata:
     w_clip_val      [    : idx1]
     w_clip_valn     [idx1: idx2]
@@ -194,19 +194,19 @@ def quant_dequant_activ(
     """
     if activ_quant_type == "per_tensor_symm":
         scale_x = 127 / a_cv
-        x_int = torch.round(x / sq * scale_x).clamp(-127, 127)
-        return x_int / scale_x * sq
+        x_int = torch.round(x / sq * scale_x).clamp(-127, 127).to(torch.int8)
+        return x_int.div(scale_x).mul(sq)
     if activ_quant_type == "per_tensor_asymm":
         scale_x = 255 / (a_cv - a_cvn)
         zp_x = a_cvn * scale_x
         x_int = torch.round(x / sq * scale_x - zp_x).clamp(0, 255)
-        return (x_int + zp_x) / scale_x * sq
+        return x_int.add(zp_x).div(scale_x).mul(sq)
     if activ_quant_type == "per_token":
         x_sq = x / sq
         a_cv_per_token = x_sq.abs().max(dim=-1, keepdim=True)[0]
         scale_x = 127 / a_cv_per_token
         x_int = torch.round(x_sq * scale_x).clamp(-127, 127)
-        return x_int / scale_x * sq
+        return x_int.div(scale_x).mul(sq)
     raise NotImplementedError(
         f"activation quantizantion type {activ_quant_type} is not supported"
     )

pyproject.toml

@@ -26,7 +26,7 @@ dependencies = [
 "accelerate>=0.20.3,!=0.34,<1.4",
 "transformers>=4.45,<4.49",
 "torch>=2.2.0,<2.5", 
-"triton>=3.0,<3.2", 
+"triton>=3.0,<3.2",
 "tqdm>=4.66.2,<5.0",
 "datasets>=3.0.0,<4.0",
 "ninja>=1.11.1.1,<2.0",

tests/aiu_addons/conftest.py

@@ -13,6 +13,9 @@
 # limitations under the License.
 """Pytest configuration file with fixtures for add-ons functionality testing"""
 
+# Standard
+from pathlib import Path
+
 # Third Party
 import pytest
 import torch
@@ -67,65 +70,58 @@ def get_gptq_gemm_inputs(request) -> tuple[torch.Tensor, ...]:
 
 i8i8_metadata = [
     {
-        "bs": 4,
-        "seq_len": 7,
-        "hid_dim": 256,
-        "out_feat": 512,
-        "dtype": torch.float16,
         "wtype": "per_tensor",  # per_channel
         "atype": "per_tensor_symm",  # per_tensor_asymm, per_token
         "smoothquant": False,
-    }
+    },
+    # {
+    #     "wtype": "per_channel",  # per_channel
+    #     "atype": "per_tensor_symm",  # per_tensor_asymm, per_token
+    #     "smoothquant": False,
+    # },
 ]
 
 
 @pytest.fixture(scope="session", params=i8i8_metadata)
 def get_i8i8_gemm_inputs(
     request,
-) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, str, str, bool]:
+) -> tuple[
+    torch.Tensor,
+    torch.Tensor,
+    torch.Tensor,
+    torch.Tensor,
+    str,
+    str,
+    bool,
+    torch.Tensor,
+]:
     """pytest fixture returning test inputs for INT8xINT8 op"""
 
     data = request.param
-    x = torch.randn(
-        (data["bs"], data["seq_len"], data["hid_dim"]),
-        dtype=data["dtype"],
-    ).clamp(-1, 1)
-    w_int = torch.randint(
-        low=-8,
-        high=8,
-        size=(data["out_feat"], data["hid_dim"]),
-        dtype=torch.int8,
+
+    filename = (
+        f"ref_w-{data['wtype']}_"
+        f"a-{data['atype']}_"
+        f"sq-{'Y' if data['smoothquant'] else 'N'}.pt"
     )
-    b = torch.zeros(data["out_feat"], dtype=data["dtype"])
-    qdata = create_qdata(
-        data["wtype"],
-        data["atype"],
-        data["hid_dim"],
-        data["out_feat"],
-        data["smoothquant"],
-        data["dtype"],
+    addon_references = Path("tests/artifacts/aiu_addons")
+    i8i8_data = torch.load(addon_references / filename, weights_only=True)
+
+    assert isinstance(i8i8_data, dict)
+    assert data["wtype"] == i8i8_data["weight_quant_type"]
+    assert data["atype"] == i8i8_data["activ_quant_type"]
+    assert data["smoothquant"] == i8i8_data["smoothquant"]
+    assert all(
+        item in i8i8_data for item in ["x", "w_int", "bias", "qdata", "reference_out"]
     )
 
-    return (x, w_int, b, qdata, data["wtype"], data["atype"], data["smoothquant"])
-
-
-def create_qdata(
-    wtype: str,
-    atype: str,
-    in_feat: int,
-    out_feat: int,
-    smoothquant: bool,
-    dtype: torch.dtype,
-) -> torch.Tensor:
-    """Generate dummy qdata tensor based on the provided quantization configuration"""
-
-    qdata_len = 2 if wtype == "per_tensor" else 2 * out_feat  # weight clips
-    qdata_len += 2  # activation clips
-    qdata_len += out_feat if atype == "per_tensor_asymm" else 1  # zero shift
-    qdata_len += in_feat if smoothquant else 1  # smoothquant scales
-
-    # TODO: improve dummy generation
-    qdata = torch.ones(qdata_len, dtype=dtype)
-    qdata[1] = -qdata[0]  # !!! temporary solution to enforce clip symmetry
-    qdata[3] = -qdata[2]
-    return qdata
+    return (
+        i8i8_data["x"],
+        i8i8_data["w_int"],
+        i8i8_data["bias"],
+        i8i8_data["qdata"],
+        i8i8_data["weight_quant_type"],
+        i8i8_data["activ_quant_type"],
+        i8i8_data["smoothquant"],
+        i8i8_data["reference_out"],
+    )

tests/aiu_addons/test_int8_addon.py

@@ -33,10 +33,17 @@ def test_i8i8_registration() -> None:
 
 def test_i8i8_op(
     get_i8i8_gemm_inputs: tuple[
-        torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, str, str, bool
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        torch.Tensor,
+        str,
+        str,
+        bool,
+        torch.Tensor,
     ],
 ) -> None:
-    """Validate output shapes of INT8xINT8 matmul.
+    """Validate output shapes and content of INT8xINT8 matmul.
     Computations are simulated, using quantized/dequantized tensors.
     """
 
@@ -48,8 +55,13 @@ def test_i8i8_op(
         weight_quant_type,
         activ_quant_type,
         smoothquant,
+        reference_out,
     ) = get_i8i8_gemm_inputs
 
+    # enforce fp16 dtype on all fp parameters for this test
+    x = x.to(torch.float16)
+    qdata = qdata.to(torch.float16)
+
     out = torch.ops.fms_mo.i8i8_aiu(
         x,
         weight,
@@ -60,4 +72,7 @@ def test_i8i8_op(
         smoothquant,
     )
 
-    assert out.size() == torch.Size((x.size()[:-1] + (weight.size(0),)))
+    error_tolerance = 1e-4  # TODO: this needs adjusting
+    assert out.size() == x.size()[:-1] + (weight.size(0),)
+    assert torch.all((out - reference_out).abs() < error_tolerance)
+    # assert torch.linalg.norm(out - reference_out) < error_tolerance  # alternative check