Merge pull request #200 from foundation-model-stack/fp8_cpu

chichun-charlie-liu · web-flow · commit 90fc888ddaf5 · 2026-03-20T15:23:57.000-04:00
fix: FP8 fallback for AIU addons running on CPU
diff --git a/fms_mo/aiu_addons/fp8/fp8_linear.py b/fms_mo/aiu_addons/fp8/fp8_linear.py
@@ -14,9 +14,11 @@
 """Implement FP8 linear module to be loaded via FMS."""
 
 # Standard
+from importlib.metadata import version
 from typing import Any, Mapping
 
 # Third Party
+from packaging.version import Version
 import torch
 
 # Local
@@ -27,6 +29,9 @@
 # torch.nn.functional.linear not recognized as callable
 # open issue in PyLint: https://github.com/pytorch/pytorch/issues/119482
 
+TORCH_VERSION = Version(torch.__version__.split("+")[0])
+SUPPORTS_CPU_PER_CHANNEL_FP8 = Version("2.10") > TORCH_VERSION
+
 # Gated torchao imports for FP8 implementation
 if available_packages["fms"] and available_packages["torchao"]:
     # Third Party
@@ -213,7 +218,11 @@ def _construct_qweight_structure(self) -> "AffineQuantizedTensor":
 
         def forward(self, x: torch.Tensor) -> torch.Tensor:
             """If input quantization is active, compute FP8xFP8 addmm leveraging torchao
-            functionalities. Otherwise compute non-quantized addmm."""
+            functionalities. Otherwise compute non-quantized addmm.
+
+            In Pytorch 2.10, torch._scale_mm only supports FP8 on CPU when
+            quantization is per-tensor. In this case, we perform a mock FP8xFP8 matmul.
+            """
             # fp8 weight tensor for torchao
             qweight: AffineQuantizedTensor = self._construct_qweight_structure()
 
@@ -234,6 +243,30 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
                     )
                 qx = self._input_activation_quant_func_fp8(x, **input_quant_kwargs)
 
+                # Check if we need CPU fallback for per-channel quantization
+                is_cpu = qx.device.type == "cpu"
+                is_per_tensor = (
+                    self.linear_config["weights"]["strategy"] == "tensor"
+                    and self.linear_config["input_activations"]["strategy"] == "tensor"
+                )
+
+                # Perform mock FP8xFP8 matmul
+                if is_cpu and not is_per_tensor and not SUPPORTS_CPU_PER_CHANNEL_FP8:
+                    # Check torchao version without loading the full package
+                    if Version("0.11") < Version(version("torchao")):
+                        raise NotImplementedError(
+                            "Fallback path for FP8 matmul on CPU is not supported "
+                            "on torchao > 0.11."
+                        )
+                    x_dequant = qx.dequantize()
+                    w_dequant = qweight.dequantize()
+                    out = torch.nn.functional.linear(
+                        x_dequant.to(w_dequant.dtype),
+                        w_dequant,
+                        self.bias if self.has_bias else None,
+                    )
+                    return out.to(x.dtype)
+
                 # Copied from torchao _linear_fp8_act_fp8_weight_impl
                 # (with changes to support fp8 out)
                 scaled_mm_config = Float8MMConfig(use_fast_accum=True)
@@ -276,10 +309,11 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
                 ).reshape(out_shape)
 
             # activations not quantized, dequant fp8 weight and do regular matmul
+            w_dequant = qweight.dequantize()
             out = torch.nn.functional.linear(
-                x, qweight.dequantize(), self.bias if self.has_bias else None
+                x.to(w_dequant.dtype), w_dequant, self.bias if self.has_bias else None
             )
-            return out
+            return out.to(x.dtype)
 
         def __repr__(self) -> str:
             return (
diff --git a/pyproject.toml b/pyproject.toml
@@ -35,7 +35,7 @@ dependencies = [
 
 [project.optional-dependencies]
 examples = ["ninja>=1.11.1.1,<2.0", "evaluate", "huggingface_hub"]
-fp8 = ["llmcompressor", "torchao==0.11"]
+fp8 = ["llmcompressor", "torchao==0.11"]  # FP8 matmul on CPU needs a fix before advancing torchao > 0.11
 gptq = ["Cython", "gptqmodel>=1.7.3"]
 mx = ["microxcaling>=1.1"]
 opt = ["fms-model-optimizer[fp8, gptq, mx]"]
diff --git a/tests/aiu_addons/test_fp8_addon.py b/tests/aiu_addons/test_fp8_addon.py
@@ -21,6 +21,84 @@
 from fms_mo.prep import available_packages
 import fms_mo.aiu_addons.fp8.fp8_spyre_op  # pylint: disable=unused-import
 
+# ============================================================================
+# Constants
+# ============================================================================
+
+# FP8 E4M3 maximum value
+FP8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+# ============================================================================
+# Helper Functions
+# ============================================================================
+
+
+def initialize_fp8_weights(
+    fp8_linear,
+    weight_strategy: str,
+    in_features: int,
+    out_features: int,
+) -> None:
+    """Initialize FP8Linear weights with proper absmax scaling.
+
+    Args:
+        fp8_linear: FP8Linear module to initialize
+        weight_strategy: "tensor" or "channel" for weight quantization
+        in_features: Input feature dimension
+        out_features: Output feature dimension
+    """
+    with torch.no_grad():
+        # Create random float weights
+        float_weights = torch.randn(out_features, in_features)
+
+        # Set appropriate scales based on strategy using absmax
+        if weight_strategy == "tensor":
+            # Per-tensor: single scale for entire weight matrix
+            absmax = float_weights.abs().max()
+            scale = absmax / FP8_E4M3_MAX
+            # Ensure scale is not zero
+            scale = torch.clamp(scale, min=1e-12)
+            fp8_linear.weight_scale.fill_(scale.item())
+        else:  # channel (per-row for weight matrix)
+            # Per-channel: one scale per output channel (row)
+            absmax = float_weights.abs().amax(dim=1)
+            scale = absmax / FP8_E4M3_MAX
+            # Ensure scales are not zero
+            scale = torch.clamp(scale, min=1e-12)
+            # Reshape to match weight_scale parameter shape (out_features, 1)
+            fp8_linear.weight_scale.copy_(scale.reshape(-1, 1))
+
+        # Quantize weights to FP8
+        quantized_weights = (float_weights / fp8_linear.weight_scale).clamp(
+            -FP8_E4M3_MAX, FP8_E4M3_MAX
+        )
+        fp8_linear.weight.copy_(quantized_weights.to(torch.float8_e4m3fn))
+
+        # Initialize bias if present
+        if fp8_linear.has_bias:
+            fp8_linear.bias.copy_(torch.randn(out_features))
+
+
+# ============================================================================
+# Pytest Fixtures
+# ============================================================================
+
+
+@pytest.fixture
+def fp8_test_dimensions():
+    """Common test dimensions for FP8Linear tests."""
+    return {
+        "batch_size": 2,
+        "seq_len": 4,
+        "in_features": 8,
+        "out_features": 16,
+    }
+
+
+# ============================================================================
+# Tests
+# ============================================================================
+
 
 def test_fp8_registration() -> None:
     """
@@ -44,9 +122,10 @@ def test_fp8_registration() -> None:
     reason="FP8 is only available on GPUs with device level 8.9 or higher",
 )
 def test_fp8_op() -> None:
-    """Validate output shapes of GPTQ W4A16 tensors.
-    Note: this AIU-compatible operation only returns a zero tensor of the
-    expected shape, it does not perform a real W4A16 matmul operation.
+    """Validate output shapes of FP8 attention operation.
+
+    Tests the FP8 attention compute operation to ensure it produces
+    outputs with the expected shape.
     """
     # Local
     from fms_mo.aiu_addons.fp8.fp8_attn import _math_fp8_compute_op
@@ -57,3 +136,140 @@ def test_fp8_op() -> None:
 
     out = _math_fp8_compute_op(query, key, value, 32, 32, 0.0, None)
     assert out.size() == query.size()
+
+
+@pytest.mark.skipif(
+    not available_packages["torchao"] or not available_packages["fms"],
+    reason="FMS and torchao required to run this test",
+)
+@pytest.mark.parametrize(
+    "weight_strategy,activation_strategy",
+    [
+        ("tensor", "tensor"),  # Per-tensor W + per-tensor dynamic A
+        ("tensor", "token"),  # Per-tensor W + per-token dynamic A
+        ("channel", "tensor"),  # Per-channel W + per-tensor dynamic A
+        ("channel", "token"),  # Per-channel W + per-token dynamic A
+    ],
+)
+def test_fp8_linear_cpu_support(  # pylint: disable=redefined-outer-name
+    weight_strategy: str,
+    activation_strategy: str,
+    fp8_test_dimensions: dict,
+) -> None:
+    """Test FP8Linear on CPU with different quantization strategies.
+
+    This test ensures that FP8Linear works correctly on CPU with:
+    - Per-tensor quantization (native support in PyTorch 2.10+)
+    - Per-channel/per-token quantization (uses fallback path in PyTorch 2.10+)
+
+    Note: PyTorch 2.10+ only supports per-tensor FP8 matmul on CPU. Per-channel
+    and per-token quantization require a fallback to dequantize + regular matmul.
+
+    Args:
+        weight_strategy: "tensor" or "channel" weight quantization
+        activation_strategy: "tensor" or "token" dynamic activation quantization
+        fp8_test_dimensions: Test dimensions fixture
+    """
+    # Local
+    from fms_mo.aiu_addons.fp8.fp8_linear import FP8Linear
+
+    # Get test dimensions
+    batch_size = fp8_test_dimensions["batch_size"]
+    seq_len = fp8_test_dimensions["seq_len"]
+    in_features = fp8_test_dimensions["in_features"]
+    out_features = fp8_test_dimensions["out_features"]
+
+    # Create FP8Linear configuration
+    linear_config = {
+        "weights": {
+            "strategy": weight_strategy,
+            "symmetric": True,
+            "dynamic": False,
+        },
+        "input_activations": {
+            "strategy": activation_strategy,
+            "symmetric": True,
+            "dynamic": True,
+        },
+    }
+
+    # Create FP8Linear module
+    fp8_linear = FP8Linear(
+        in_features=in_features,
+        out_features=out_features,
+        bias=True,
+        linear_config=linear_config,
+    )
+
+    # Initialize weights using helper function
+    initialize_fp8_weights(fp8_linear, weight_strategy, in_features, out_features)
+
+    # Create input tensor on CPU
+    x = torch.randn(batch_size, seq_len, in_features, dtype=torch.bfloat16)
+
+    # Run forward pass - should not raise an error
+    output = fp8_linear(x)
+
+    # Validate output shape
+    assert output.shape == (batch_size, seq_len, out_features)
+
+    # Validate output is not NaN or Inf
+    assert not torch.isnan(output).any()
+    assert not torch.isinf(output).any()
+
+    # Validate output dtype matches input dtype
+    assert output.dtype == x.dtype
+
+
+@pytest.mark.skipif(
+    not available_packages["torchao"] or not available_packages["fms"],
+    reason="FMS and torchao required to run this test",
+)
+def test_fp8_linear_cpu_no_activation_quantization(fp8_test_dimensions: dict) -> None:  # pylint: disable=redefined-outer-name
+    """Test FP8Linear on CPU with only weight quantization (no activation quantization).
+
+    This tests the code path where activations are not quantized but weights are FP8.
+
+    Args:
+        fp8_test_dimensions: Test dimensions fixture
+    """
+    # Local
+    from fms_mo.aiu_addons.fp8.fp8_linear import FP8Linear
+
+    # Get test dimensions
+    batch_size = fp8_test_dimensions["batch_size"]
+    seq_len = fp8_test_dimensions["seq_len"]
+    in_features = fp8_test_dimensions["in_features"]
+    out_features = fp8_test_dimensions["out_features"]
+
+    # Create FP8Linear configuration with no activation quantization
+    linear_config = {
+        "weights": {
+            "strategy": "channel",
+            "symmetric": True,
+            "dynamic": False,
+        },
+        "input_activations": None,  # No activation quantization
+    }
+
+    # Create FP8Linear module
+    fp8_linear = FP8Linear(
+        in_features=in_features,
+        out_features=out_features,
+        bias=True,
+        linear_config=linear_config,
+    )
+
+    # Initialize weights using helper function
+    initialize_fp8_weights(fp8_linear, "channel", in_features, out_features)
+
+    # Create input tensor on CPU
+    x = torch.randn(batch_size, seq_len, in_features, dtype=torch.bfloat16)
+
+    # Run forward pass
+    output = fp8_linear(x)
+
+    # Validate output
+    assert output.shape == (batch_size, seq_len, out_features)
+    assert not torch.isnan(output).any()
+    assert not torch.isinf(output).any()
