NVFP4 quantization emulation kernels as reference (#4324)

Summary:
Pull Request resolved: #4324

X-link: facebookresearch/FBGEMM#1397

Add the NVFP4 quantization emulation kernels as reference.
These kernels were provided by Nvidia and used for the FP4 QAT study on LLaMa3 8B.
We can use these kernels for numerical studies on H100.

Reviewed By: jiawenliu64

Differential Revision: D76363519

fbshipit-source-id: e3341d1ef38591eae061d1a457a1c1e82b336ebb

Author: summerdengfb

fbgemm_gpu/experimental/gen_ai/src/quantize/quantize.cpp

@@ -262,6 +262,12 @@ void scaled_fp4_quant(
     at::Tensor const& output_sf,
     at::Tensor const& input_sf);
 
+std::vector<at::Tensor> fake_quantize_nvfp4_per_tensor(
+    at::Tensor input,
+    std::optional<at::Tensor> static_scales,
+    std::optional<at::Tensor> bs, // batch size
+    std::optional<at::Tensor> scale_ub); // scale upperbound
+
 TORCH_LIBRARY_IMPL(fbgemm, CUDA, m) {
   m.impl("f8f8bf16_blockwise", f8f8bf16_blockwise);
   m.impl("f8f8bf16_tensorwise", f8f8bf16_tensorwise);
@@ -308,6 +314,7 @@ TORCH_LIBRARY_IMPL(fbgemm, CUDA, m) {
   m.impl("bf16i4bf16_rowwise", bf16i4bf16_rowwise);
   m.impl("scaled_fp4_quant", scaled_fp4_quant);
   m.impl("i8i8bf16_dynamic", i8i8bf16_dynamic);
+  m.impl("fake_quantize_nvfp4_per_tensor", fake_quantize_nvfp4_per_tensor);
 #endif
 
 #ifdef USE_ROCM
@@ -357,6 +364,7 @@ TORCH_LIBRARY_IMPL(fbgemm, CPU, m) {
   m.impl("bf16i4bf16_rowwise_batched", bf16i4bf16_rowwise_batched);
   m.impl("bf16i4bf16_rowwise", bf16i4bf16_rowwise);
   m.impl("scaled_fp4_quant", scaled_fp4_quant);
+  m.impl("fake_quantize_nvfp4_per_tensor", fake_quantize_nvfp4_per_tensor);
 #endif
 }