bitsandbytes, test existence of triton.ops.matmul_perf_model

This fixes the bitsandbytes to work with newer triton versions
which does not anymore include the triton.ops.matmul_perf_model
which has been moved to triton-langs kernels project.
https://github.com/triton-lang/kernels

Fix is to simply catch the import error if triton is available
but the triton.ops.matmul_perf_model is not. In such case the logic
is implemented in a same way than earlier when the triton
was not available at all.

To my understanding there is currently no a way to install
opt_matmul_perf.py from new tritons kernel project directly with pip,
so in future it could propably be good to include that python class
directly in bitsandbytes.

bitsandbytes-foundation#1492

Signed-off-by: Mika Laitio <[email protected]>

bitsandbytes/triton/int8_matmul_mixed_dequantize.py

@@ -2,7 +2,16 @@
 
 from bitsandbytes.triton.triton_utils import is_triton_available
 
-if not is_triton_available():
+use_triton = False
+if is_triton_available():
+    import triton
+    try:
+        from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+        use_triton = True
+    except ImportError:
+        print("bitsandbytes matmul_mixed warning, triton.ops.matmul_perf_model is not available anymore.")
+
+if use_triton == False:
 
     def int8_matmul_mixed_dequantize(a, b, state_x, state_w, bias):
         return None

bitsandbytes/triton/int8_matmul_rowwise_dequantize.py

@@ -2,7 +2,16 @@
 
 from bitsandbytes.triton.triton_utils import is_triton_available
 
-if not is_triton_available():
+use_triton = False
+if is_triton_available():
+    import triton
+    try:
+        from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
+        use_triton = True
+    except ImportError:
+        print("bitsandbytes rowwise_dequantize warning, triton.ops.matmul_perf_model is not available anymore.")
+
+if use_triton == False:
 
     def int8_matmul_rowwise_dequantize(a, b, state_x, state_w, bias):
         return None