[Blockwise] Add support block_len=64 support

Add env to chose blocklen of blockwise quantize.
Signed-off-by: wenjh <wenjh@sugon.com>

Fix pytest of blockwise error
Signed-off-by: wenjh <wenjh@sugon.com>

Resolve new api in  int8 gemm test
Signed-off-by: wenjh <wenjh@sugon.com>

Fix incorrect launch parm
Signed-off-by: wenjh <wenjh@sugon.com>

Fix 1D blockwise(64) acc error
Signed-off-by: wenjh <wenjh@sugon.com>

transformer_engine/pytorch/triton/blockwise_int8_gemm_nt.py

@@ -8,6 +8,7 @@ import triton
 import triton.language as tl
 import pandas as pd
 from transformer_engine.pytorch.triton.per_token_group_quant import _int8_gemm_helper
+from transformer_engine.pytorch.fp8 import blockwise_fp8_block_len
 import functools
 import logging
 
@@ -557,7 +558,7 @@ def apply_w8a8_block_int8_linear_batched_helper(m: int,
                                             k: int,
                                             out_dtype: Type[torch.dtype] = torch.float16,
                                             device: str = "cuda",
-                                            block_size: List[int]=[128,128],
+                                            block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                                             bias: Optional[torch.Tensor] = None,
                                             best_config:Optional[dict]=None):
     
@@ -596,7 +597,7 @@ def apply_w8a8_block_int8_linear_helper(m: int,
                                             k: int,
                                             out_dtype: Type[torch.dtype] = torch.float16,
                                             device: str = "cuda",
-                                            block_size: List[int]=[128,128],
+                                            block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                                             bias: Optional[torch.Tensor] = None,
                                             best_config:Optional[dict]=None):
     
@@ -771,7 +772,7 @@ def main():
     n_list=[7168]
     k_list=[1152]
 
-    block_size=[128, 128]
+    block_size=[blockwise_fp8_block_len, blockwise_fp8_block_len]
     
     out_dtype=torch.bfloat16
     

transformer_engine/pytorch/triton/blockwise_int8_gemm_nt_wgrad.py

@@ -8,6 +8,7 @@ import triton
 import triton.language as tl
 import pandas as pd
 from transformer_engine.pytorch.triton.per_token_group_quant import _int8_gemm_helper_b
+from transformer_engine.pytorch.fp8 import blockwise_fp8_block_len
 import functools
 import logging
 
@@ -596,7 +597,7 @@ def apply_w8a8_block_int8_linear_batched_helper(m: int,
                                             k: int,
                                             out_dtype: Type[torch.dtype] = torch.float16,
                                             device: str = "cuda",
-                                            block_size: List[int]=[128,128],
+                                            block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                                             bias: Optional[torch.Tensor] = None,
                                             best_config:Optional[dict]=None):
     
@@ -639,7 +640,7 @@ def apply_w8a8_block_int8_linear_helper(m: int,
                                             k: int,
                                             out_dtype: Type[torch.dtype] = torch.float16,
                                             device: str = "cuda",
-                                            block_size: List[int]=[128,128],
+                                            block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                                             bias: Optional[torch.Tensor] = None,
                                             best_config:Optional[dict]=None):
     
@@ -821,7 +822,7 @@ def main():
     n_list=[7168]
     k_list=[1152]
 
-    block_size=[128, 128]
+    block_size=[blockwise_fp8_block_len, blockwise_fp8_block_len]
     
     out_dtype=torch.bfloat16
     

transformer_engine/pytorch/triton/per_token_group_quant.py

@@ -9,6 +9,7 @@ import triton.language as tl
 import pandas as pd
 import logging
 import math
+from transformer_engine.pytorch.fp8 import blockwise_fp8_block_len
 
 
 def to_int8(tensor: torch.Tensor):
@@ -118,7 +119,7 @@ def _int8_gemm_helper(m: int,
                              k: int,
                              out_dtype: Type[torch.dtype] = torch.float16,
                              device: str = "cuda",
-                             block_size: List[int]=[128,128],
+                             block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                              best_config:Optional[list] = None):
     # Test for a cutlass kernel with per-token activation quantization
     # and per-output channel weight quantization.
@@ -143,7 +144,7 @@ def _int8_gemm_helper_b(m: int,
                              k: int,
                              out_dtype: Type[torch.dtype] = torch.float16,
                              device: str = "cuda",
-                             block_size: List[int]=[128,128],
+                             block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                              best_config:Optional[list] = None):
     # Test for a cutlass kernel with per-token activation quantization
     # and per-output channel weight quantization.
@@ -168,7 +169,7 @@ def _int8_gemm_helper_test(m: int,
                              k: int,
                              out_dtype: Type[torch.dtype] = torch.float16,
                              device: str = "cuda",
-                             block_size: List[int]=[128,128],
+                             block_size: List[int]=[blockwise_fp8_block_len,blockwise_fp8_block_len],
                              best_config:Optional[list] = None):
     # Test for a cutlass kernel with per-token activation quantization
     # and per-output channel weight quantization.
@@ -197,7 +198,7 @@ def main():
     m_list=[1,2,4,8,16,32,64,128,256,512,1024,2048,4096,8192,16384,32768]
     n_list=[576,2048,7168,256,7168,1536,1536]
     k_list=[7168,512,1024,7168,128,7168,1536]
-    block_size=[128,128]
+    block_size=[blockwise_fp8_block_len,blockwise_fp8_block_len]
     out_dtype=torch.bfloat16
     _n=[]
     _k=[]