Merge branch 'master' into mrwyattii/pydantic-2-support

microsoft · May 28, 2024 · 62ca5f2 · 62ca5f2
2 parents aee5f9d + 4deb40d
commit 62ca5f2
Show file tree

Hide file tree

Showing 6 changed files with 7 additions and 5 deletions.
diff --git a/csrc/fp_quantizer/quantize.cpp → csrc/fp_quantizer/fp_quantize.cpp b/csrc/fp_quantizer/quantize.cpp → csrc/fp_quantizer/fp_quantize.cpp
@@ -3,7 +3,7 @@
 
 // DeepSpeed Team
 
-#include "quantize.h"
+#include "fp_quantize.h"
 
 #include <c10/cuda/CUDAStream.h>
 #include <torch/extension.h>

diff --git a/csrc/fp_quantizer/quantize.cu → csrc/fp_quantizer/fp_quantize.cu b/csrc/fp_quantizer/quantize.cu → csrc/fp_quantizer/fp_quantize.cu
@@ -5,8 +5,8 @@
 
 #include <stdexcept>
 #include "context.h"
+#include "fp_quantize.h"
 #include "memory_access_utils.h"
-#include "quantize.h"
 #include "reduction_utils.h"
 
 #include <cuda.h>

diff --git a/csrc/fp_quantizer/includes/quantize.h → csrc/fp_quantizer/includes/fp_quantize.h b/csrc/fp_quantizer/includes/quantize.h → csrc/fp_quantizer/includes/fp_quantize.h
diff --git a/deepspeed/runtime/zero/mics.py b/deepspeed/runtime/zero/mics.py
@@ -43,7 +43,7 @@ def wait(self) -> None:
         """
         # let the current stream to op
         try:
-            print("HANDLE", self.allgather_handle)
+            # print("HANDLE", self.allgather_handle)
             instrument_w_nvtx(self.allgather_handle.wait)()
         except (ValueError, RuntimeError) as e:
             log_dist(

diff --git a/docs/_sass/minimal-mistakes/_sidebar.scss b/docs/_sass/minimal-mistakes/_sidebar.scss
@@ -79,6 +79,7 @@
     top: auto;
     right: 0;
     width: $right-sidebar-width-narrow;
+    margin-right: -1.5 * $right-sidebar-width-narrow;
     padding-left: 1em;
     z-index: 10;
 
@@ -93,6 +94,7 @@
 
   @include breakpoint($x-large) {
     width: $right-sidebar-width;
+    margin-right: -1.5 * $right-sidebar-width-narrow;
   }
 }
 

diff --git a/op_builder/fp_quantizer.py b/op_builder/fp_quantizer.py
@@ -52,8 +52,8 @@ def filter_ccs(self, ccs):
 
     def sources(self):
         return [
-            "csrc/fp_quantizer/quantize.cu",
-            "csrc/fp_quantizer/quantize.cpp",
+            "csrc/fp_quantizer/fp_quantize.cu",
+            "csrc/fp_quantizer/fp_quantize.cpp",
         ]
 
     def extra_ldflags(self):