[float8] improve eager numerics for dynamic scales and gets on par with torch.compile (#904)

* [float8] improve eager numerics for dynamic scales

* leave torch.linalg.vector_norm for another PR

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* cuda

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* remove _data and investigate

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* remove _data comment

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* upcast to float32 is enough

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* explain why float32

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* _data parity

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* handle sm8.9

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* fix transformer unit test

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* print if error

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* Add tutorial for trainable tensor subclass (#908)

Summary: The new tutorial provides an example of how to implement
a trainable tensor subclass that wraps quantized data. This extends
the existing `MyDTypeTensor` with a few necessary steps to ensure
proper gradient updates, namely:

1. Define a differentiable constructor
2. Define backward pass for ops of interest (e.g. torch.nn.functional.linear)
3. Handle special ops used by the optimizer (e.g. aten.add, aten.add_)

Test Plan:
python tutorials/developer_api_guide/my_trainable_tensor_subclass.py

* Introducing 1-bit quantization for Llama in torchchat (#910)

Differential Revision: D63052325

Pull Request resolved: #911

* Rename Floating point to fp8 (#909)

* [float8] fix typo in bitwise_identical unit test (#918)

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* Adding example for quantized tensor + tensor parallelism (#785)

* [WIP] Adding example for quantized tensor + tensor parallelism

Summary:
This PR adds an example of how quantized tensor subclass can work with DTensor: https://github.com/pytorch/pytorch/blob/main/torch/distributed/_tensor/README.md

End goal is to rewrite https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/llama2.py with normal llama2 implementation and show case with DTensor + AffineQuantizedTensor + torch.compile we can get on par performance with the custom tensor parallel implementation

Test Plan:
torchrun --standalone --nnodes=1 --nproc-per-node=4 tutorials/developer_api_guide/tensor_parallel.py

Reviewers:

Subscribers:

Tasks:

Tags:

* tensor parallel file

* Use DTensor.from instead of distribute_tensor

* implementing aten.slice.Tensor (WIP)

* working

* some shape fix and use more quant primitive ops

* Add rowwise test

* make rowwise sharding work

* compile still not working yet

* fake tensor didn't pick up shape changes from transpose

* backend='eager'

* change transpose to non-inplace op

* add error message

* works now with torch nightly

* remove print

* ruff

* Clean up

* Fix device id

---------

Co-authored-by: Ke Wen <[email protected]>

* rename cuda mode -> gpu mode (#925)

* Add workaround to recover the perf for quantized vit in torch.compile (#926)

Add temporary workaround to recover the perf for quantized vit under torch.compile

Summary:
Recently we found a perf drop in quantized vit due to #898 (comment)
This PR add a temp fix until we figure out the longer term fix.

I think ideally we should figure out why the tensor subclass check failed in torch.compile (https://github.com/pytorch/pytorch/blob/e4d294221b140fdbb49a64f297bc60c9fcc2f80e/torch/nn/modules/activation.py#L1286) and fix that

Test Plan:
python tutorials/quantize_vit/run_vit_b_quant.py

Reviewers:

Subscribers:

Tasks:

Tags:

* clean up device checks in float8 unit test files (#923)

Summary:

While working on rowwise scaling I noticed that some of the CUDA
device capability checks we had in the test files did not make sense,
cleaning this up.

Test Plan:

tests pass on my H100

CI, it should skip less tests now since CI only has CUDA capability 8, 9

Reviewers:

Subscribers:

Tasks:

Tags:

* [low-bit optim] Change 8-bit and FP8 optim block size from 2048 to 256 to match new bnb v0.44 (#927)

* Float8 autoquant weight only (#866)

* Fix failing FP6 benchmark (#931)

* Remove two if statements in fp8 padding (#935)

Reviewed By: vkuzo

Differential Revision: D63051205

Pull Request resolved: #935
Approved by: https://github.com/vkuzo

* [Distributed] Improve sharding example (#937)

* [Distributed] Improve sharding example

* Add comment

* Add composable QAT quantizer (#938)

Summary: This is a utility for users who wish to apply multiple
QAT quantizers to their models. In the near future, we expect
to add an embedding QAT quantizer that composes with the
existing linear QAT quantizers.

Test Plan:
python test/quantization/test_qat.py -k test_composable_qat_quantizer

* resolve conflict with latest main

Differential Revision: D63048850

Pull Request resolved: #912

* Add torchchat quantizer

Differential Revision: D62394341

Pull Request resolved: #897

* Add compile tests to test suite (#906)

* Add compile tests to test suite

Summary:
This is a follow up PR addressing #839 (comment)
We can add more compiler related tests in the future.

Next
* refactor a bit to use quantize_ API directly
* use the test suite in existing API tests

Test Plan:
python torchao/testing/utils.py

Reviewers:

Subscribers:

Tasks:

Tags:

* rename

* add result check

* Fix up CMakeLists and reorganize some code locations

Differential Revision: D62711903

Pull Request resolved: #948

* [float8] all-reduce amax on dp mesh instead of global pg (#933)

* [float8] all-reduce amax on dp mesh instead of global pg

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* liner

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* improve comments

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* move hp tensor inside if

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* linter

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* linter

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* linter

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* linter

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* linter

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* int8 dynamic quant + bsr support (#821)

This PR, adds in int8 dynamicquant + bsr support.

Changes:
* Use i8i8 -> bf16 matmul to maintain accuracy
* Added a block sparse layout type to AffineQuantizedTensor + check/impl.  
* Cleaned up benchmark.py script and add a single line `benchmark.sh` file for acceleration numbers
* Updated eval.py and added a single line `evaluate.sh` file for accuracy numbers
* Lots of lint formatting and README updates
* torch.compile now working and is correct

* fixing some issues with our support for 70/405B models (#941)

Summary: download and convert scripts needed to be updated alongside
model.py config files

Test Plan: python generate.py --checkpoint_path ../../../checkpoints/meta-llama/Meta-Llama-3.1-70B/model.pth

Reviewers:

Subscribers:

Tasks:

Tags:

* Update INT8 mixed-precision training test to be less flaky (#950)

* Add executorch parallel

Differential Revision: D62711909

Pull Request resolved: #953

* test CI

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* better comment on why upcasting

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* control seed

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* move unit test to test_compile

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* fix typo

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* float64 upcasting after allreduce

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

* use LinearMMConfig

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

---------

Co-authored-by: andrewor14 <[email protected]>
Co-authored-by: Vaishnavi Gupta <[email protected]>
Co-authored-by: Apurva Jain <[email protected]>
Co-authored-by: Jerry Zhang <[email protected]>
Co-authored-by: Ke Wen <[email protected]>
Co-authored-by: Mark Saroufim <[email protected]>
Co-authored-by: Vasiliy Kuznetsov <[email protected]>
Co-authored-by: Thien Tran <[email protected]>
Co-authored-by: Tobias van der Werff <[email protected]>
Co-authored-by: Shuqi Yang <[email protected]>
Co-authored-by: Scott Roy <[email protected]>
Co-authored-by: Jesse Cai <[email protected]>
Co-authored-by: HDCharles <[email protected]>

Author: 14 people

test/float8/test_compile.py

@@ -25,8 +25,15 @@
     get_float8_layers,
     sync_float8_amax_and_scale_history,
 )
-from torchao.float8.float8_scaling_utils import hp_tensor_to_float8_delayed
-from torchao.float8.float8_tensor import LinearMMConfig
+from torchao.float8.float8_scaling_utils import (
+    hp_tensor_to_float8_delayed,
+    hp_tensor_to_float8_dynamic,
+)
+from torchao.float8.float8_tensor import (
+    LinearMMConfig,
+    GemmInputRole,
+    ScaledMMConfig,
+)
 from torchao.float8.float8_utils import e4m3_dtype
 
 from torch._dynamo.test_case import TestCase as DynamoTestCase
@@ -353,5 +360,65 @@ def test_sync_amax_func_cuda_graph_success():
     assert "skipping cudagraphs due to mutaton on input" not in stderr[0]
 
 
+@unittest.skipIf(
+        not is_cuda_8_9,
+        "CUDA not available",
+    )
+@pytest.mark.parametrize(
+    "dtype",
+    [
+        torch.float32,
+        torch.bfloat16,
+        torch.float16,
+    ],
+)
+def test_dynamic_scale_numeric_parity(dtype: torch.dtype):
+    scaling_type_weight = ScalingType.DYNAMIC
+    torch.manual_seed(42)
+    hp_tensor1 = torch.randn(16, 16, device="cuda", dtype=dtype)
+    hp_tensor2 = hp_tensor1.detach().clone()
+    float8_config = Float8LinearConfig(
+        cast_config_weight=CastConfig(scaling_type=scaling_type_weight),
+    )
+    linear_mm_config = LinearMMConfig(
+        # output
+        ScaledMMConfig(
+            False,
+            float8_config.gemm_config_output.use_fast_accum,
+            False,
+            float8_config.pad_inner_dim,
+        ),
+        # grad_input
+        ScaledMMConfig(
+            False,
+            float8_config.gemm_config_grad_input.use_fast_accum,
+            False,
+            float8_config.pad_inner_dim,
+        ),
+        # grad_weight
+        ScaledMMConfig(
+            False,
+            float8_config.gemm_config_grad_weight.use_fast_accum,
+            False,
+            float8_config.pad_inner_dim,
+        ),
+    )
+    float8_eager = hp_tensor_to_float8_dynamic(
+        hp_tensor1,
+        torch.float8_e4m3fn,
+        linear_mm_config,
+        gemm_input_role=GemmInputRole.WEIGHT,
+    )
+    torch._dynamo.reset()
+    float8_compile = torch.compile(hp_tensor_to_float8_dynamic)(
+        hp_tensor2,
+        torch.float8_e4m3fn,
+        linear_mm_config,
+        gemm_input_role=GemmInputRole.WEIGHT,
+    )
+    assert torch.equal(float8_eager._scale, float8_compile._scale)
+    assert torch.equal(float8_eager._data, float8_compile._data)
+
+
 if __name__ == "__main__":
     pytest.main([__file__])

torchao/float8/float8_tensor.py

@@ -163,7 +163,10 @@ def forward(
 
         DTensor Invariant: DTensor must always be the outer most tensor subclass
         """
-        tensor_scaled = tensor * scale
+        # Note: when the line below is compiled with `torch.compile`, `tensor` is automatically
+        # upcasted to `float32` to multiply with the scale
+        # In order to match numerics between eager and compile, we upcast manually here.
+        tensor_scaled = tensor.to(torch.float32) * scale
         bits_fp8 = to_fp8_saturated(tensor_scaled, float8_dtype)
 
         if isinstance(bits_fp8, DTensor):

torchao/float8/float8_utils.py

@@ -42,6 +42,9 @@ def amax_to_scale(
         float8_dtype: The float8 dtype.
         orig_dtype: The original dtype of the tensor.
     """
+    # torch.compile and eager show different numerics for 1.0 / float32,
+    # upcast to float64 to ensure same numeric between compile and eager
+    amax = amax.to(torch.float64)
     if float8_dtype in FP8_TYPES:
         res = torch.finfo(float8_dtype).max / torch.clamp(amax, min=EPS)
     else:

torchao/float8/fsdp_utils.py

@@ -64,12 +64,17 @@ def precompute_float8_dynamic_scale_for_fsdp(module: nn.Module) -> None:
     # clamp is dispatched through DTensor
     # it will issue a single all-reduce
     amax_tensor = torch.clamp(amax_tensor, EPS)  # Replicate
+    # keep consistent with float8_utils.amax_to_scale
+    # torch.compile and eager show different numerics for 1.0 / float32,
+    # upcast to float64 to ensure same numeric between compile and eager
+    origin_dtype = amax_tensor.dtype
+    amax_tensor = amax_tensor.to(torch.float64)
     scale_tensor = torch.finfo(torch.float8_e4m3fn).max / amax_tensor  # Replicate
-    if amax_tensor.dtype is torch.float16:
+    if origin_dtype is torch.float16:
         scale_tensor = torch.clamp(scale_tensor, max=torch.finfo(torch.float16).max)
-    local_scale_tensor = scale_tensor.to_local()
+    local_scale_tensor = scale_tensor.to_local().to(torch.float32)
     for i, float8_linear in enumerate(float8_linears):
-        float8_linear.weight._local_tensor._precomputed_scale = local_scale_tensor[i].to(torch.float32)
+        float8_linear.weight._local_tensor._precomputed_scale = local_scale_tensor[i]
 
 
 # FSDP pads its local tensor on dim-0. The subclass should be preserved such

torchao/testing/float8/fsdp2_utils.py

@@ -48,10 +48,7 @@ def check_parity_no_mp(
             ):
                 precompute_float8_dynamic_scale_for_fsdp(model)
 
-        if compile_transformer_block:
-            test_cls.assertEqual(losses[0], losses[1], atol=1e-4, rtol=1e-4, msg = f"iter: {iter_idx}, loss-ref: {losses[0]}, loss-fp8: {losses[1]}")
-        else:
-            test_cls.assertEqual(losses[0], losses[1], msg = f"iter: {iter_idx}, loss-ref: {losses[0]}, loss-fp8: {losses[1]}")
+        test_cls.assertEqual(losses[0], losses[1], msg = f"iter: {iter_idx}, loss-ref: {losses[0]}, loss-fp8: {losses[1]}")
 
 
 def check_parity_bf16_mp(