[DCU] Skip some tests in test_sanity.py

Signed-off-by: zhaochao <zhaochao1@sugon.com>

tests/pytorch/test_sanity.py

@@ -46,7 +46,7 @@ from utils import ModelConfig
 
 # Only run FP8 tests on supported devices.
 fp8_available, reason_for_no_fp8 = FP8GlobalStateManager.is_fp8_available()
-fp8_block_scaling_available, _ = FP8GlobalStateManager.is_fp8_block_scaling_available()
+fp8_block_scaling_available, reason_for_no_fp8_block_scaling = FP8GlobalStateManager.is_fp8_block_scaling_available()
 mxfp8_available, reason_for_no_mxfp8 = FP8GlobalStateManager.is_mxfp8_available()
 
 # Record initial RNG state from script run.
@@ -378,7 +378,13 @@ def test_sanity_layernorm_linear(
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -436,7 +442,13 @@ def test_sanity_linear_with_zero_tokens(dtype, bs, model, fp8_recipe, fp8_model_
     num_tokens = bs * config.max_seqlen_q
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     use_fp8 = fp8_recipe is not None
@@ -525,7 +537,13 @@ def test_sanity_layernorm_mlp(
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -567,7 +585,13 @@ def test_sanity_gpt(
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -733,7 +757,13 @@ def test_sanity_amp_and_nvfuser(dtype, fp8_recipe, model, skip_wgrad):
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -763,7 +793,13 @@ def test_sanity_drop_path(dtype, fp8_recipe, model):
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -797,7 +833,13 @@ def test_sanity_fused_qkv_params(dtype, fp8_recipe, model, skip_wgrad):
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023
@@ -831,7 +873,13 @@ def test_sanity_gradient_accumulation_fusion(dtype, fp8_recipe, model, skip_wgra
     config = model_configs[model]
 
     if fp8_recipe is not None:
-        if not is_fp8_supported(config):
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if fp8_recipe.float8_block_scaling() and not fp8_block_scaling_available:
+            pytest.skip(reason_for_no_fp8_block_scaling)
+        if fp8_recipe.mxfp8() and not mxfp8_available:
+            pytest.skip(reason_for_no_mxfp8)
+        if not config.is_fp8_supported():
             pytest.skip("Model config does not support FP8")
 
     sigma = 0.023