[Others] support import deepgemm/deepep from fleet ops (#6351)

* update paddleformers to v1.0

* only change import fleetpath

fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py

@@ -28,24 +28,23 @@ from fastdeploy.platforms import current_platform
 from fastdeploy.utils import register_custom_python_op
 from fastdeploy.worker.tbo import let_another_thread_run
 
-from ..utils import get_sm_version
 from .fused_moe_backend_base import MoEMethodBase
 from .fused_moe_triton_backend import BlockWiseFP8MoEMethod
 
 if current_platform.is_cuda():
-    if get_sm_version() == 100:
-        logger.info("Detected sm100, use PFCC DeepGEMM")
-        paddle.compat.enable_torch_proxy(scope={"deep_gemm"})
-        from deep_gemm import (
-            m_grouped_fp8_gemm_nt_contiguous,
-            m_grouped_fp8_gemm_nt_masked,
+    try:
+        m_grouped_fp8_gemm_nt_contiguous = (
+            fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.m_grouped_fp8_gemm_nt_contiguous
         )
-    else:
-        from fastdeploy.model_executor.ops.gpu.deep_gemm import (
-            m_grouped_gemm_fp8_fp8_bf16_nt_contiguous as m_grouped_fp8_gemm_nt_contiguous,
+        m_grouped_fp8_gemm_nt_masked = (
+            fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.m_grouped_fp8_gemm_nt_masked
         )
-        from fastdeploy.model_executor.ops.gpu.deep_gemm import (
-            m_grouped_gemm_fp8_fp8_bf16_nt_masked as m_grouped_fp8_gemm_nt_masked,
+    except:
+        m_grouped_fp8_gemm_nt_contiguous = (
+            fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous
+        )
+        m_grouped_fp8_gemm_nt_masked = (
+            fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_masked
         )
 else:
     m_grouped_fp8_gemm_nt_contiguous = None

fastdeploy/model_executor/layers/quantization/block_wise_fp8.py

@@ -42,14 +42,10 @@ from ..utils import get_sm_version, get_tensor, per_block_cast_to_fp8
 from .quant_base import QuantConfigBase, QuantMethodBase
 
 if current_platform.is_cuda():
-    if get_sm_version() == 100:
-        # SM100 should use PFCC DeepGemm
-        paddle.compat.enable_torch_proxy(scope={"deep_gemm"})
-        from deep_gemm import fp8_gemm_nt
-    else:
-        from fastdeploy.model_executor.ops.gpu.deep_gemm import (
-            gemm_fp8_fp8_bf16_nt as fp8_gemm_nt,
-        )
+    try:
+        fp8_gemm_nt = fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.fp8_gemm_nt
+    except:
+        fp8_gemm_nt = fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.gemm_fp8_fp8_bf16_nt
 else:
     fp8_gemm_nt = None
 

fastdeploy/model_executor/layers/quantization/fp8_utils.py

@@ -21,16 +21,36 @@ from fastdeploy.platforms import current_platform
 
 from ..utils import get_sm_version
 
-if current_platform.is_cuda():
-    if get_sm_version() == 100:
-        # SM100 should use PFCC DeepGemm
-        logger.info("Detected sm100, use PFCC DeepGEMM")
-        paddle.compat.enable_torch_proxy(scope={"deep_gemm"})
-        import deep_gemm
+
+def load_deep_gemm():
+    """
+    Load DeepGemm module according to FastDeploy env switch.
+
+    Returns:
+        Imported deep_gemm module object.
+    """
+
+    if current_platform.is_cuda():
+        if get_sm_version() == 100:
+            # SM100 should use PFCC DeepGemm
+            paddle.compat.enable_torch_proxy(scope={"deep_gemm"})
+            try:
+                from paddlefleet.ops import deep_gemm
+
+                logger.info("Detected sm100, use PaddleFleet DeepGEMM")
+            except:
+                import deep_gemm
+
+                logger.info("Detected sm100, use PFCC DeepGEMM")
+        else:
+            logger.info("use FastDeploy DeepGEMM")
+            from fastdeploy.model_executor.ops.gpu import deep_gemm
     else:
-        from fastdeploy.model_executor.ops.gpu import deep_gemm
-else:
-    deep_gemm = None
+        deep_gemm = None
+    return deep_gemm
+
+
+deep_gemm = load_deep_gemm()
 
 
 def ceil_div(x: int, y: int) -> int:

fastdeploy/model_executor/layers/utils.py

@@ -23,6 +23,7 @@ from paddle import Tensor, nn
 from paddle.framework import in_dynamic_mode
 from scipy.linalg import block_diag
 
+import fastdeploy
 from fastdeploy.config import FDConfig
 from fastdeploy.platforms import current_platform
 
@@ -254,10 +255,7 @@ def per_block_cast_to_fp8(x: Tensor, block_size: list = [128, 128]) -> Tuple[Ten
     Only used in deep_gemm block wise quant weight.
     copy from FastDeploy/custom_ops/gpu_ops/fp8_deep_gemm/tests/test_core.py.
     """
-    try:
-        from deep_gemm import ceil_div
-    except ModuleNotFoundError:
-        from fastdeploy.model_executor.ops.gpu.deep_gemm import ceil_div
+    ceil_div = fastdeploy.model_executor.layers.quantization.fp8_utils.deep_gemm.ceil_div
 
     assert x.dim() == 2
     m, n = x.shape