[MRG] relax the FP8 CUDA arch limitation to SM89 (#549)

closes: #548

> Nvidia Ada Lovelace GPUs (e.g., RTX 4090, L20, L40) with SM89 version
are also support FP8 MMA, and hence, it is recommended to relax the CUDA
architecture limitations to enable FP8 training on a broader range of
devices.
>
> and the [CUDA 12.0
announcement](https://developer.nvidia.com/blog/cuda-toolkit-12-0-released-for-general-availability/)
says that it supports Lovelace architecture:
> '*CUDA 12.0 exposes programmable functionality for many features of
the NVIDIA Hopper and NVIDIA Ada Lovelace architectures: ...32x Ultra
xMMA (including FP8 and FP16)*'
> 
> -
https://developer.nvidia.com/blog/cuda-toolkit-12-0-released-for-general-availability/
> - https://nvidia.github.io/TensorRT-LLM/reference/support-matrix.html
> -
https://github.com/NVIDIA/cutlass/blob/c4e3e122e266644c61b4af33d0cc09f4c391a64b/include/cutlass/arch/mma_sm89.h#L57
> 
>
![image](https://github.com/user-attachments/assets/3c11736c-2e84-4bd6-a49c-5af8b0e3e6ac)

After relaxing the CUDA architecture limitations for FP8, my environment
with **4 x L40 GPUs (SM89)** can still successfully train llama under
float8 precision.


![image](https://github.com/user-attachments/assets/1337e041-0d0d-49b5-8c11-00e67f4df41f)

---------

Co-authored-by: Andrew Gu <[email protected]>

torchtitan/float8.py

@@ -23,9 +23,9 @@
 from torchtitan.parallelisms import ParallelDims
 
 
-def _is_sm90_or_later():
-    # Float8 is only supported on H100+ GPUs
-    return torch.cuda.is_available() and torch.cuda.get_device_capability() >= (9, 0)
+def _is_sm89_or_later():
+    # Float8 is only supported on SM89 or later (H100+ GPUs)
+    return torch.cuda.is_available() and torch.cuda.get_device_capability() >= (8, 9)
 
 
 class Float8Handler:
@@ -35,9 +35,9 @@ def __init__(self, job_config: JobConfig, parallel_dims: ParallelDims):
         float8_config = job_config.float8
         if not float8_config.enable_float8_linear:
             return
-        if not _is_sm90_or_later():
+        if not _is_sm89_or_later():
             logger.warning(
-                "Failed to swap to Float8Linear because SM90 or later is not available",
+                "Failed to swap to Float8Linear because float8 is only supported on SM89 or later",
             )
             return
         try: