Merge branch 'master' into Tecorigin-SDAA-accelerator

.github/workflows/cpu-torch-latest.yml

@@ -42,7 +42,7 @@ jobs:
           git clone https://github.com/huggingface/transformers
           cd transformers
           # if needed switch to the last known good SHA until transformers@master is fixed
-          git checkout 6c3f168b3
+          # git checkout 6c3f168b3
           git rev-parse --short HEAD
           pip install .
 

.github/workflows/no-torch.yml

@@ -4,6 +4,7 @@ on:
   workflow_dispatch:
   pull_request:
     paths:
+      - 'accelerator/**'
       - '.github/workflows/no-torch.yml'
       - 'op_builder/**'
   schedule:

.github/workflows/nv-torch-latest-v100.yml

@@ -38,7 +38,7 @@ jobs:
           git clone https://github.com/huggingface/transformers
           cd transformers
           # if needed switch to the last known good SHA until transformers@master is fixed
-          git checkout 6c3f168b3
+          # git checkout 6c3f168b3
           git rev-parse --short HEAD
           pip install .
 

.github/workflows/nv-torch-nightly-v100.yml

@@ -2,6 +2,9 @@ name: nv-torch-nightly-v100
 
 on:
   workflow_dispatch:
+    pull_request:
+    paths:
+      - '.github/workflows/nv-torch-nightly-v100.yml'
   schedule:
     - cron: "0 0 * * *"
 
@@ -34,7 +37,7 @@ jobs:
           git clone https://github.com/huggingface/transformers
           cd transformers
           # if needed switch to the last known good SHA until transformers@master is fixed
-          # git checkout 1cc453d33
+          # git checkout 6c3f168b3
           git rev-parse --short HEAD
           pip install .
 

.github/workflows/xpu-compile.yml

@@ -31,10 +31,10 @@ jobs:
       run: |
         apt-get update
         apt-get install clinfo libaio-dev python3-pip -y
-        pip install torch==2.3.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/torch/
-        pip install intel-extension-for-pytorch==2.3.110+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/intel-extension-for-pytorch/
-        pip install oneccl_bind_pt==2.3.100+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/oneccl-bind-pt/
-        pip install torchvision==0.18.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/torchvision/
+        pip install torch==2.3.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/torch/
+        pip install intel-extension-for-pytorch==2.3.110+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/intel-extension-for-pytorch/
+        pip install oneccl_bind_pt==2.3.100+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/oneccl-bind-pt/
+        pip install torchvision==0.18.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/torchvision/
         pip install https://github.com/intel/intel-xpu-backend-for-triton/releases/download/v3.0.0b2/triton_xpu-3.0.0b2-cp310-cp310-linux_x86_64.whl
         pip install py-cpuinfo numpy
         pip install .[dev,autotuning]

.github/workflows/xpu-max1100.yml

@@ -47,10 +47,10 @@ jobs:
       run: |
         apt-get update
         apt-get install clinfo libaio-dev python3-pip -y
-        pip install torch==2.3.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/torch/
-        pip install intel-extension-for-pytorch==2.3.110+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/intel-extension-for-pytorch/
-        pip install oneccl_bind_pt==2.3.100+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/oneccl-bind-pt/
-        pip install torchvision==0.18.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/us/torchvision/
+        pip install torch==2.3.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/torch/
+        pip install intel-extension-for-pytorch==2.3.110+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/intel-extension-for-pytorch/
+        pip install oneccl_bind_pt==2.3.100+xpu -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/oneccl-bind-pt/
+        pip install torchvision==0.18.1 -f https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/torchvision/
         pip install py-cpuinfo numpy
         pip install .[dev,autotuning]
 

CODEOWNERS

@@ -8,49 +8,52 @@
 
 # top-level repo folders
 /.github/ @loadams
-/azure/ @awan-10
-/benchmarks/ @awan-10 @tjruwase
+/azure/ @loadams
+/benchmarks/ @guanhuawang @tjruwase
 /bin/ @loadams
-/csrc/ @awan-10
+/csrc/ @tjruwase
 /deepspeed/ @loadams @tjruwase
-/docker/ @awan-10
+/docker/ @loadams @guanhuawang
 /docs/ @loadams @tjruwase
-/examples/ @awan-10 @tohtana
+/examples/ @jomayeri @tohtana
 /op_builder/ @loadams @tjruwase @jomayeri
-/release/ @loadams
+/release/ @loadams @jomayeri
 /requirements/ @loadams
-/scripts/ @awan-10
+/scripts/ @loadams @tjruwase
 /tests/ @tjruwase @loadams @tohtana
 
 # deepspeed
 /deepspeed/autotuning/ @loadams
 /deepspeed/checkpoint/ @tjruwase
-/deepspeed/comm/ @awan-10
+/deepspeed/comm/ @guanhuawang
 /deepspeed/compression/ @tjruwase
-/deepspeed/elasticity/ @awan-10
+/deepspeed/elasticity/ @tjruwase
 /deepspeed/launcher/ @loadams
-/deepspeed/module_inject/ @awan-10
+/deepspeed/module_inject/ @hwchen2017 @loadams
 /deepspeed/moe/ @tohtana
-/deepspeed/monitor/ @awan-10
+/deepspeed/monitor/ @tjruwase
 /deepspeed/nebula/ @tjruwase
+/deepspeed/nvme/ @tjruwase @jomayeri
 /deepspeed/ops/ @tohtana
 /deepspeed/pipe/ @tohtana @loadams
 /deepspeed/profiling/ @loadams
-/deepspeed/utils/ @tjruwase @awan-10
+/deepspeed/sequence/ @tohtana
+/deepspeed/utils/ @tjruwase @tohtana
 
 # inference
-/deepspeed/inference/ @awan-10
-/deepspeed/model_implementations/ @awan-10
+/deepspeed/inference/ @hwchen2017 @tohtana
+/deepspeed/model_implementations/@tohtana @loadams
 
 # training
 /deepspeed/runtime/ @tjruwase @tohtana
 /deepspeed/runtime/activation_checkpointing/ @tjruwase
 /deepspeed/runtime/checkpoint_engine/ @tjruwase
-/deepspeed/runtime/comm/ @awan-10
-/deepspeed/runtime/compression/ @awan-10
+/deepspeed/runtime/comm/ @guanhuawang
+/deepspeed/runtime/compression/ @tjruwase
 /deepspeed/runtime/data_pipeline/ @tjruwase
-/deepspeed/runtime/fp16/ @tjruwase
-/deepspeed/runtime/fp16/onebit/ @awan-10
-/deepspeed/runtime/pipe/ @loadams
-/deepspeed/runtime/swap_tensor/ @tjruwase
-/deepspeed/runtime/zero/ @tjruwase
+/deepspeed/runtime/domino/ @guanhuawang @hwchen2017
+/deepspeed/runtime/fp16/ @tjruwase @tohtana
+/deepspeed/runtime/fp16/onebit/ @tjruwase
+/deepspeed/runtime/pipe/ @loadams @tohtana
+/deepspeed/runtime/swap_tensor/ @tjruwase @jomayeri
+/deepspeed/runtime/zero/ @tjruwase @tohtana

COMMITTERS.md

@@ -5,5 +5,7 @@
 | Olatunji Ruwase | [tjruwase](https://github.com/tjruwase)     | Microsoft |
 | Logan Adams     | [loadams](https://github.com/loadams)      | Microsoft |
 | Masahiro Tanaka | [tohtana](https://github.com/tohtana)      | Microsoft |
-| Jeff Rasley     | [jeffra](https://github.com/jeffra)       | SnowFlake |
-| Minjia Zhang    | [minjiazhang](https://github.com/minjiazhang)  | UIUC      |
+| Jeff Rasley     | [jeffra](https://github.com/jeffra)       | SnowFlake  |
+| Minjia Zhang    | [minjiazhang](https://github.com/minjiazhang)  | UIUC  |
+| Ashwin Aji      | [ashwinma](https://github.com/ashwinma)        | AMD   |
+| Sam Foreman     | [saforem2](https://github.com/saforem2)        | Argonne National Laboratory |

README.md

@@ -27,7 +27,7 @@
 * [2023/11] [DeepSpeed ZeRO-Offload++: 6x Higher Training Throughput via Collaborative CPU/GPU Twin-Flow](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-offloadpp)
 * [2023/11] [DeepSpeed-FastGen: High-throughput Text Generation for LLMs via MII and DeepSpeed-Inference](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-fastgen) [[English](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-fastgen)] [[中文](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-fastgen/chinese/README.md)] [[日本語](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-fastgen/japanese/README.md)]
 * [2023/10] [DeepSpeed-VisualChat: Improve Your Chat Experience with Multi-Round Multi-Image Inputs](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-visualchat/10-03-2023/README.md) [[English](https://github.com/microsoft/DeepSpeed/tree/master/blogs/deepspeed-visualchat/10-03-2023/README.md)] [[中文](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed-visualchat/10-03-2023/README-Chinese.md)] [[日本語](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed-visualchat/10-03-2023/README-Japanese.md)]
-* [2023/09] Announcing the DeepSpeed4Science Initiative: Enabling large-scale scientific discovery through sophisticated AI system technologies [[DeepSpeed4Science website](https://deepspeed4science.ai/)] [[Tutorials](https://www.deepspeed.ai/deepspeed4science/)] [[White paper](https://arxiv.org/abs/2310.04610)] [[Blog](https://www.microsoft.com/en-us/research/blog/announcing-the-deepspeed4science-initiative-enabling-large-scale-scientific-discovery-through-sophisticated-ai-system-technologies/)] [[中文](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed4science/chinese/README.md)] [[日本語](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed4science/japanese/README.md)]
+* [2023/09] Announcing the DeepSpeed4Science Initiative: Enabling large-scale scientific discovery through sophisticated AI system technologies [[Tutorials](https://www.deepspeed.ai/deepspeed4science/)] [[White paper](https://arxiv.org/abs/2310.04610)] [[Blog](https://www.microsoft.com/en-us/research/blog/announcing-the-deepspeed4science-initiative-enabling-large-scale-scientific-discovery-through-sophisticated-ai-system-technologies/)] [[中文](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed4science/chinese/README.md)] [[日本語](https://github.com/microsoft/DeepSpeed/blob/master/blogs/deepspeed4science/japanese/README.md)]
 
 
 <!-- NOTE: we must use html for news items otherwise links will be broken in the 'more news' section -->
@@ -78,7 +78,7 @@ To further increase the inference efficiency, DeepSpeed offers easy-to-use and f
 
 ## DeepSpeed4Science
 
-In line with Microsoft's mission to solve humanity's most pressing challenges, the DeepSpeed team at Microsoft is responding to this opportunity by launching a new initiative called *DeepSpeed4Science*, aiming to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. Learn more: [DeepSpeed4Science website](https://deepspeed4science.ai/) and [tutorials](https://www.deepspeed.ai/deepspeed4science/)
+In line with Microsoft's mission to solve humanity's most pressing challenges, the DeepSpeed team at Microsoft is responding to this opportunity by launching a new initiative called *DeepSpeed4Science*, aiming to build unique capabilities through AI system technology innovations to help domain experts to unlock today's biggest science mysteries. Learn more: [tutorials](https://www.deepspeed.ai/deepspeed4science/)
 
 ---
 

accelerator/cpu_accelerator.py

@@ -3,9 +3,15 @@
 
 # DeepSpeed Team
 
-import torch
 from .abstract_accelerator import DeepSpeedAccelerator
 
+# During setup stage torch may not be installed, pass on no torch will
+# allow op builder related API to be executed.
+try:
+    import torch
+except ImportError as e:
+    pass
+
 try:
     import oneccl_bindings_for_pytorch  # noqa: F401 # type: ignore
     oneccl_imported_p = True

accelerator/hpu_accelerator.py

@@ -21,8 +21,8 @@ def __init__(self):
         self.apply_hpu_workarounds()
         try:
             import habana_frameworks.torch.hpu as hpu
-            hpu.setDeterministic(True)
             self.hpu = hpu
+            torch.use_deterministic_algorithms(True)
         except ImportError as e:
             raise ValueError(
                 f"HPU_Accelerator requires habana_frameworks.torch.hpu, which is not installed on this system.")

accelerator/real_accelerator.py

@@ -131,10 +131,9 @@ def get_accelerator():
         if accelerator_name is None:
             try:
                 import intel_extension_for_pytorch as ipex
+
                 if ipex._C._has_xpu():
                     accelerator_name = "xpu"
-                else:
-                    accelerator_name = "cpu"
             except ImportError as e:
                 pass
         if accelerator_name is None:
@@ -175,23 +174,27 @@ def get_accelerator():
             except ImportError as e:
                 pass
         if accelerator_name is None:
-            # borrow this log from PR#5084
             try:
                 import torch
 
                 # Determine if we are on a GPU or x86 CPU with torch.
-                if torch.cuda.is_available():  #ignore-cuda
+                # "torch.cuda.is_available()" provides a stronger guarantee,     #ignore-cuda
+                # ensuring that we are free from CUDA initialization errors.
+                # While "torch.cuda.device_count() > 0" check ensures that       #ignore-cuda
+                # we won't try to do any CUDA calls when no device is available
+                # For reference: https://github.com/microsoft/DeepSpeed/pull/6810
+                if torch.cuda.device_count() > 0 and torch.cuda.is_available():  #ignore-cuda
                     accelerator_name = "cuda"
-                else:
-                    if accel_logger is not None:
-                        accel_logger.warn(
-                            "Setting accelerator to CPU. If you have GPU or other accelerator, we were unable to detect it."
-                        )
-                    accelerator_name = "cpu"
             except (RuntimeError, ImportError) as e:
                 # TODO need a more decent way to detect which accelerator to use, consider using nvidia-smi command for detection
-                accelerator_name = "cuda"
                 pass
+        if accelerator_name is None:
+            # borrow this log from PR#5084
+            if accel_logger is not None:
+                accel_logger.warn(
+                    "Setting accelerator to CPU. If you have GPU or other accelerator, we were unable to detect it.")
+            # cpu added as catch-all when accelerator detection fails
+            accelerator_name = "cpu"
 
         ds_set_method = "auto detect"
 

csrc/includes/quantization.h

@@ -52,6 +52,36 @@ void launch_swizzled_quant(int8_t* q_data,
                            int devices_per_node,
                            cudaStream_t stream);
 
+void launch_loco_swizzled_quant(int8_t* quantized_data,
+                                float* quantized_scales,
+                                const __half* uncompressed_data,
+                                __half* error_feedback,
+                                const float err_beta,
+                                int num_bits,
+                                quantize::Type quant_type,
+                                int groups,
+                                int elems_per_group,
+                                int pipelining,
+                                int nodes,
+                                int devices_per_node,
+                                cudaStream_t stream);
+
+void launch_loco_dequant_reduce(int8_t* reduced_data,
+                                float* reduced_scales,
+                                const int8_t* input_data,
+                                const float* input_scales,
+                                int num_gpus,
+                                int num_bits,
+                                quantize::Type quant_type,
+                                int out_groups,
+                                int elems_per_out_group,
+                                int elems_per_in_tensor,
+                                int groups_per_in_tensor,
+                                int elems_per_in_group,
+                                __half2* error_feedback,
+                                const float err_beta,
+                                cudaStream_t stream);
+
 void launch_dequant_reduce(int8_t* reduced_data,
                            float* reduced_scales,
                            const int8_t* input_data,

csrc/includes/quantization_utils.h

@@ -24,6 +24,7 @@ constexpr int max_threads = 1024;
 Class to hold the quantization parameters for a given tensor.
 Holds the implementation of the quantization operation.
 */
+
 template <Type qType, int numBits>
 class Params {
 public: