merge main into multi-backend-refactor

.github/workflows/build_documentation.yml

@@ -13,7 +13,9 @@ jobs:
     with:
       commit_sha: ${{ github.sha }}
       package: bitsandbytes
-      repo_owner: TimDettmers
+      repo_owner: bitsandbytes-foundation
+      # avoid /src suffix leading to wrong links, like bitsandbytes/blob/main/src/bitsandbytes/nn/
+      version_tag_suffix: ''  # defaults to '/src'
       custom_container: huggingface/transformers-doc-builder
     secrets:
       hf_token: ${{ secrets.HUGGINGFACE_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -9,11 +9,13 @@ concurrency:
 
 jobs:
   build:
-    if: github.repository == 'TimDettmers/bitsandbytes'
+    if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
     with:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}
       package: bitsandbytes
-      repo_owner: TimDettmers
+      repo_owner: bitsandbytes-foundation
+      # avoid /src suffix leading to wrong links, like bitsandbytes/blob/main/src/bitsandbytes/nn/
+      version_tag_suffix: ''  # defaults to '/src'
       custom_container: huggingface/transformers-doc-builder

.github/workflows/python-package.yml

@@ -63,12 +63,10 @@ jobs:
         os: [ubuntu-latest, windows-latest]
         arch: [x86_64, aarch64]
         cuda_version:
-          ["11.7.1", "11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.0"]
+          ["11.7.1", "11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.0"]
         exclude:
           - os: windows-latest # This probably requires arm64 Windows agents
             arch: aarch64
-          - os: windows-latest  # The Jimver/cuda-toolkit is action used for Windows builds is not updated for 12.4 yet.
-            cuda_version: "12.4.0"
           - os: ubuntu-latest # Temporary. Takes too long, not ready yet.
             arch: aarch64
     runs-on: ${{ matrix.os }} # One day, we could run them on native agents. Azure supports this now but it's planned only for Q3 2023 for hosted agents
@@ -79,7 +77,7 @@ jobs:
         if: startsWith(matrix.os, 'ubuntu')
         uses: docker/setup-qemu-action@v2
         # Windows: We install Cuda on the agent (slow)
-      - uses: Jimver/[email protected].14
+      - uses: Jimver/[email protected].16
         if: startsWith(matrix.os, 'windows')
         id: cuda-toolkit
         with:

CHANGELOG.md

@@ -1,3 +1,29 @@
+### 0.43.2
+
+This release is quite significant as the QLoRA bug fix big implications for higher `seqlen` and batch sizes.
+
+For each sequence (i.e. batch size increase of one) we expect memory savings of:
+- 405B: 39GB for seqlen 1024, and 4888GB for 128k
+- 70B: 20.1GB for 1024 and 2516GB for 128k
+
+This was due to activations being unnecessary for frozen parameters, yet the memory for them was still erroneously allocated due to the now fixed bug.
+
+#### Improvements:
+
+- docs: FSDP+QLoRA and CPU install guide (#1211 #1227, thanks @stevhliu)
+- Add CUDA 12.5 and update 12.4 builds (#1284)
+
+#### Bug Fixes
+
+- 4bit getstate and 8bit deepcopy (#1230 #1231, thanks @BenjaminBossan)
+- missing optimizers in `str2optimizer32bit` (#1222, thanks @EtienneDosSantos)
+- CUDA 12.5 build issue (#1273, thanks @HennerM)
+- fix for min_8bit_size functionality in Optimizer base classes (#1286, thanks @Edenzzzz)
+- QLoRA mem bug (#1270, thanks @Ther-nullptr)
+- tests for cpu only platforms (#1259, thanks @galqiwi)
+- restoration of quant_storage for CPU offloading (#1279)
+- optim update error with non-contiguous grads/params (deepspeed) (#1187)
+
 ### 0.43.1
 
 #### Improvements:

CMakeLists.txt

@@ -77,6 +77,13 @@ endif()
 
 
 if(BUILD_CUDA)
+    # NVCC normally will only work with MSVC up to 1939. VS2022 17.10+ starts using versions 1940+.
+    # Workaround: use --allow-unsupported-compiler
+    # This needs to be added *before* we try to enable the CUDA language so CMake's compiler check passes.
+    if(MSVC AND MSVC_VERSION VERSION_GREATER_EQUAL 1940)
+        string(APPEND CMAKE_CUDA_FLAGS " --allow-unsupported-compiler")
+    endif()
+
     enable_language(CUDA) # This will fail if CUDA is not found
     find_package(CUDAToolkit REQUIRED)
 
@@ -229,7 +236,6 @@ if(WIN32)
     set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
 endif()
 
-# Weird MSVC hacks
 if(MSVC)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2 /fp:fast")
 endif()

_typos.toml

@@ -1,5 +1,10 @@
 [files]
 
+[default]
+extend-ignore-re = [
+    "@Ther-nul",  # valid Github user
+]
+
 [default.extend-identifiers]
 
 [type.py.extend-words]

bitsandbytes/__init__.py

@@ -74,4 +74,4 @@
     "optim.optimizer.MockArgs": False,
 }
 
-__version__ = "0.43.2.dev"
+__version__ = "0.43.3.dev"

bitsandbytes/autograd/_functions.py

@@ -524,7 +524,7 @@ def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState]
         ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype
 
         if any(ctx.needs_input_grad[:2]):
-            ctx.tensors = (A, B)
+            ctx.tensors = (None, B)
         else:
             ctx.tensors = (None, None)
 
@@ -537,7 +537,7 @@ def backward(ctx, grad_output):
             return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
 
         req_gradA, _, _, req_gradBias, _ = ctx.needs_input_grad
-        A, B = ctx.tensors
+        _, B = ctx.tensors
 
         grad_A, grad_B, grad_bias = None, None, None
 

bitsandbytes/functional.py

@@ -27,6 +27,35 @@ def prod(iterable):
 
 if lib and lib.compiled_with_cuda:
     """C FUNCTIONS FOR OPTIMIZERS"""
+    str2optimizer32bit = {
+        "adam": (
+            lib.cadam32bit_grad_fp32,
+            lib.cadam32bit_grad_fp16,
+            lib.cadam32bit_grad_bf16,
+        ),
+        "momentum": (
+            lib.cmomentum32bit_grad_32,
+            lib.cmomentum32bit_grad_16,
+        ),
+        "rmsprop": (
+            lib.crmsprop32bit_grad_32,
+            lib.crmsprop32bit_grad_16,
+        ),
+        "lion": (
+            lib.clion32bit_grad_fp32,
+            lib.clion32bit_grad_fp16,
+            lib.clion32bit_grad_bf16,
+        ),
+        "adagrad": (
+            lib.cadagrad32bit_grad_32,
+            lib.cadagrad32bit_grad_16,
+        ),
+        "lamb": (
+            lib.cadam32bit_grad_fp32,
+            lib.cadam32bit_grad_fp16,
+        ),
+    }
+
     str2optimizer8bit = {
         "adam": (
             lib.cadam_static_8bit_grad_32,

bitsandbytes/nn/modules.py

@@ -240,7 +240,7 @@ def __new__(
         return self
 
     def __getstate__(self):
-        state = self.__dict__
+        state = self.__dict__.copy()
         state["data"] = self.data
         state["requires_grad"] = self.requires_grad
         return state
@@ -286,6 +286,9 @@ def from_prequantized(
         self.compress_statistics = self.quant_state.nested
         self.quant_type = self.quant_state.quant_type
         self.bnb_quantized = True
+
+        self.quant_storage = data.dtype
+
         return self
 
     def _quantize(self, device):
@@ -340,6 +343,7 @@ def to(self, *args, **kwargs):
                 blocksize=self.blocksize,
                 compress_statistics=self.compress_statistics,
                 quant_type=self.quant_type,
+                quant_storage=self.quant_storage,
             )
 
             return new_param
@@ -457,7 +461,7 @@ def forward(self, x: torch.Tensor):
                 # since we registered the module, we can recover the state here
                 assert self.weight.shape[1] == 1
                 if not isinstance(self.weight, Params4bit):
-                    self.weight = Params4bit(self.weight, quant_storage=self.quant_storage)
+                    self.weight = Params4bit(self.weight, quant_storage=self.quant_storage, bnb_quantized=True)
                 self.weight.quant_state = self.quant_state
             else:
                 print(
@@ -567,13 +571,12 @@ def __new__(
         CB=None,
         SCB=None,
     ):
-        cls.has_fp16_weights = has_fp16_weights
-        cls.CB = None
-        cls.SCB = None
         if data is None:
             data = torch.empty(0)
         obj = torch.Tensor._make_subclass(cls, data, requires_grad)
-        obj.CB, obj.SCB = cls.CB, cls.SCB
+        obj.CB = CB
+        obj.SCB = SCB
+        obj.has_fp16_weights = has_fp16_weights
         return obj
 
     def cuda(self, device):
@@ -592,6 +595,18 @@ def cuda(self, device):
 
         return self
 
+    def __deepcopy__(self, memo):
+        # adjust this if new arguments are added to the constructor
+        new_instance = type(self).__new__(
+            type(self),
+            data=copy.deepcopy(self.data, memo),
+            requires_grad=self.requires_grad,
+            has_fp16_weights=self.has_fp16_weights,
+            CB=copy.deepcopy(self.CB, memo),
+            SCB=copy.deepcopy(self.SCB, memo),
+        )
+        return new_instance
+
     def cpu(self):
         # we store the 8-bit rows-major weight
         B = self.data.contiguous().bfloat16().cpu()

bitsandbytes/optim/optimizer.py

@@ -437,7 +437,7 @@ def init_state(self, group, p, gindex, pindex):
         state = self.state[p]
         state["step"] = 0
 
-        if dtype == torch.float32 or (dtype == torch.uint8 and p.numel() < 4096):
+        if dtype == torch.float32:
             state["state1"] = self.get_state_buffer(p, dtype=torch.float32)
             state["state2"] = self.get_state_buffer(p, dtype=torch.float32)
         elif dtype == torch.uint8:
@@ -474,6 +474,10 @@ def init_state(self, group, p, gindex, pindex):
 
     @torch.no_grad()
     def update_step(self, group, p, gindex, pindex):
+        # avoid update error from non-contiguous memory layout
+        p.data = p.data.contiguous()
+        p.grad = p.grad.contiguous()
+
         state = self.state[p]
         grad = p.grad
 
@@ -656,7 +660,7 @@ def init_state(self, group, p, gindex, pindex):
         state = self.state[p]
         state["step"] = 0
 
-        if dtype == torch.float32 or (dtype == torch.uint8 and p.numel() < 4096):
+        if dtype == torch.float32:
             state["state1"] = self.get_state_buffer(p, dtype=torch.float32)
         elif dtype == torch.uint8:
             if state["step"] == 0:
@@ -685,6 +689,10 @@ def init_state(self, group, p, gindex, pindex):
 
     @torch.no_grad()
     def update_step(self, group, p, gindex, pindex):
+        # avoid update error from non-contiguous memory layout
+        p.data = p.data.contiguous()
+        p.grad = p.grad.contiguous()
+
         state = self.state[p]
         grad = p.grad
 

csrc/kernels.cu

@@ -12,8 +12,6 @@
 #include <cub/block/block_reduce.cuh>
 #include <cub/cub.cuh>
 #include <math_constants.h>
-#include <thrust/host_vector.h>
-#include <thrust/device_vector.h>
 #include <mma.h>
 
 

csrc/ops.cuh

@@ -19,10 +19,6 @@
 #include <vector>
 #include <functional>
 
-#include <thrust/host_vector.h>
-#include <thrust/device_vector.h>
-
-
 
 #define CUDA_CHECK_RETURN(value) {                      \
   cudaError_t _m_cudaStat = value;                    \