Merge branch 'main' into feat/add-cu_seqlens

Author: zigzagcai

.github/workflows/publish.yaml

@@ -43,15 +43,24 @@ jobs:
           # Using ubuntu-20.04 instead of 22.04 for more compatibility (glibc). Ideally we'd use the
           # manylinux docker image, but I haven't figured out how to install CUDA on manylinux.
           os: [ubuntu-20.04]
-          python-version: ['3.7', '3.8', '3.9', '3.10', '3.11']
-          torch-version: ['1.12.1', '1.13.1', '2.0.1', '2.1.2', '2.2.0']
+          python-version: ['3.7', '3.8', '3.9', '3.10', '3.11', '3.12']
+          torch-version: ['1.12.1', '1.13.1', '2.0.1', '2.1.2', '2.2.0', '2.3.0.dev20240105']
           cuda-version: ['11.8.0', '12.2.2']
           # We need separate wheels that either uses C++11 ABI (-D_GLIBCXX_USE_CXX11_ABI) or not.
           # Pytorch wheels currently don't use it, but nvcr images have Pytorch compiled with C++11 ABI.
           # Without this we get import error (undefined symbol: _ZN3c105ErrorC2ENS_14SourceLocationESs)
           # when building without C++11 ABI and using it on nvcr images.
           cxx11_abi: ['FALSE', 'TRUE']
           exclude:
+            # Pytorch < 2.2 does not support Python 3.12
+            - torch-version: '1.12.1'
+              python-version: '3.12'
+            - torch-version: '1.13.1'
+              python-version: '3.12'
+            - torch-version: '2.0.1'
+              python-version: '3.12'
+            - torch-version: '2.1.2'
+              python-version: '3.12'
             # Pytorch <= 1.12 does not support Python 3.11
             - torch-version: '1.12.1'
               python-version: '3.11'
@@ -62,6 +71,8 @@ jobs:
               python-version: '3.7'
             - torch-version: '2.2.0'
               python-version: '3.7'
+            - torch-version: '2.3.0.dev20240105'
+              python-version: '3.7'
             # Pytorch <= 2.0 only supports CUDA <= 11.8
             - torch-version: '1.12.1'
               cuda-version: '12.2.2'
@@ -119,12 +130,24 @@ jobs:
           # If we don't install before installing Pytorch, we get error for torch 2.0.1
           # ERROR: Could not find a version that satisfies the requirement setuptools>=40.8.0 (from versions: none)
           pip install lit
+          # For some reason torch 2.2.0 on python 3.12 errors saying no setuptools
+          pip install setuptools
           # We want to figure out the CUDA version to download pytorch
           # e.g. we can have system CUDA version being 11.7 but if torch==1.12 then we need to download the wheel from cu116
           # This code is ugly, maybe there's a better way to do this.
-          export TORCH_CUDA_VERSION=$(python -c "import os; minv = {'1.12': 113, '1.13': 116, '2.0': 117, '2.1': 118, '2.2': 118}[os.environ['MATRIX_TORCH_VERSION']]; maxv = {'1.12': 116, '1.13': 117, '2.0': 118, '2.1': 121, '2.2': 121}[os.environ['MATRIX_TORCH_VERSION']]; print(max(min(int(os.environ['MATRIX_CUDA_VERSION']), maxv), minv))")
+          export TORCH_CUDA_VERSION=$(python -c "from os import environ as env; \
+            minv = {'1.12': 113, '1.13': 116, '2.0': 117, '2.1': 118, '2.2': 118, '2.3': 118}[env['MATRIX_TORCH_VERSION']]; \
+            maxv = {'1.12': 116, '1.13': 117, '2.0': 118, '2.1': 121, '2.2': 121, '2.3': 121}[env['MATRIX_TORCH_VERSION']]; \
+            print(max(min(int(env['MATRIX_CUDA_VERSION']), maxv), minv))" \
+          )
           if [[ ${{ matrix.torch-version }} == *"dev"* ]]; then
-            pip install --no-cache-dir --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
+            if [[ ${MATRIX_TORCH_VERSION} == "2.2" ]]; then
+              # --no-deps because we can't install old versions of pytorch-triton
+              pip install typing-extensions jinja2
+              pip install --no-cache-dir --no-deps --pre https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}/torch-${{ matrix.torch-version }}%2Bcu${TORCH_CUDA_VERSION}-cp${MATRIX_PYTHON_VERSION}-cp${MATRIX_PYTHON_VERSION}-linux_x86_64.whl
+            else
+              pip install --no-cache-dir --pre torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/nightly/cu${TORCH_CUDA_VERSION}
+            fi
           else
             pip install --no-cache-dir torch==${{ matrix.torch-version }} --index-url https://download.pytorch.org/whl/cu${TORCH_CUDA_VERSION}
           fi

README.md

@@ -13,7 +13,7 @@ with an efficient hardware-aware design and implementation in the spirit of [Fla
 
 ## Installation
 
-- `pip install causal-conv1d>=1.1.0,<1.2.0`: an efficient implementation of a simple causal Conv1d layer used inside the Mamba block.
+- [Option] `pip install causal-conv1d>=1.2.0`: an efficient implementation of a simple causal Conv1d layer used inside the Mamba block.
 - `pip install mamba-ssm`: the core Mamba package.
 
 It can also be built from source with `pip install .` from this repository.

mamba_ssm/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "1.1.4"
+__version__ = "1.2.0.post1"
 
 from mamba_ssm.ops.selective_scan_interface import selective_scan_fn, mamba_inner_fn
 from mamba_ssm.modules.mamba_simple import Mamba

mamba_ssm/models/config_mamba.py

@@ -12,3 +12,4 @@ class MambaConfig:
     residual_in_fp32: bool = True
     fused_add_norm: bool = True
     pad_vocab_size_multiple: int = 8
+    tie_embeddings: bool = True

mamba_ssm/models/mixer_seq_simple.py

@@ -220,8 +220,9 @@ def __init__(
         self.tie_weights()
 
     def tie_weights(self):
-        self.lm_head.weight = self.backbone.embedding.weight
-
+        if self.config.tie_embeddings:
+            self.lm_head.weight = self.backbone.embedding.weight
+    
     def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
         return self.backbone.allocate_inference_cache(batch_size, max_seqlen, dtype=dtype, **kwargs)
 

mamba_ssm/modules/mamba_simple.py

@@ -15,7 +15,7 @@
 try:
     from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
 except ImportError:
-    causal_conv1d_fn, causal_conv1d_update = None
+    causal_conv1d_fn, causal_conv1d_update = None, None
 
 try:
     from mamba_ssm.ops.triton.selective_state_update import selective_state_update
@@ -143,7 +143,7 @@ def forward(self, hidden_states, cu_seqlens=None, inference_params=None):
 
         A = -torch.exp(self.A_log.float())  # (d_inner, d_state)
         # In the backward pass we write dx and dz next to each other to avoid torch.cat
-        if self.use_fast_path and inference_params is None:  # Doesn't support outputting the states
+        if self.use_fast_path and causal_conv1d_fn is not None and inference_params is None:  # Doesn't support outputting the states
             out = mamba_inner_fn(
                 xz,
                 self.conv1d.weight,

mamba_ssm/ops/selective_scan_interface.py

@@ -6,8 +6,13 @@
 
 from einops import rearrange, repeat
 
-from causal_conv1d import causal_conv1d_fn
-import causal_conv1d_cuda
+try:
+    from causal_conv1d import causal_conv1d_fn
+    import causal_conv1d_cuda
+except ImportError:
+    causal_conv1d_fn = None
+    causal_conv1d_cuda = None
+
 import selective_scan_cuda
 
 
@@ -168,6 +173,7 @@ def forward(ctx, xz, conv1d_weight, conv1d_bias, x_proj_weight, delta_proj_weigh
         """
              xz: (batch, dim, seqlen)
         """
+        assert causal_conv1d_cuda is not None, "causal_conv1d_cuda is not available. Please install causal-conv1d."
         assert checkpoint_lvl in [0, 1]
         L = xz.shape[-1]
         delta_rank = delta_proj_weight.shape[1]
@@ -196,7 +202,7 @@ def forward(ctx, xz, conv1d_weight, conv1d_bias, x_proj_weight, delta_proj_weigh
             assert x.shape[2] == (d_conv - 1) * len(cu_seqlens[1:-1]) + z.shape[2]
 
         conv1d_bias = conv1d_bias.contiguous() if conv1d_bias is not None else None
-        conv1d_out = causal_conv1d_cuda.causal_conv1d_fwd(x, conv1d_weight, conv1d_bias, None, True)
+        conv1d_out = causal_conv1d_cuda.causal_conv1d_fwd(x, conv1d_weight, conv1d_bias, None, None, None, True)
 
         # (Optional Step2 for cu_seqlens): Mask conv1d ops in cumulative sequences
         if cu_seqlens is not None:
@@ -262,6 +268,7 @@ def forward(ctx, xz, conv1d_weight, conv1d_bias, x_proj_weight, delta_proj_weigh
     @custom_bwd
     def backward(ctx, dout):
         # dout: (batch, seqlen, dim)
+        assert causal_conv1d_cuda is not None, "causal_conv1d_cuda is not available. Please install causal-conv1d."
         (xz, conv1d_weight, conv1d_bias, x_dbl, x_proj_weight, delta_proj_weight, out_proj_weight,
          conv1d_out, delta, A, B, C, D, delta_bias, scan_intermediates, out, cu_seqlens) = ctx.saved_tensors
         L = xz.shape[-1]
@@ -285,7 +292,7 @@ def backward(ctx, dout):
                 x = padded_x
                 assert x.shape[2] == (d_conv - 1) * len(cu_seqlens[1:-1]) + z.shape[2]
 
-            conv1d_out = causal_conv1d_cuda.causal_conv1d_fwd(x, conv1d_weight, conv1d_bias, None, True)
+            conv1d_out = causal_conv1d_cuda.causal_conv1d_fwd(x, conv1d_weight, conv1d_bias, None, None, None, True)
 
             # (Optional Step2 for cu_seqlens): Mask conv1d ops in cumulative sequences
             if cu_seqlens is not None:
@@ -345,8 +352,8 @@ def backward(ctx, dout):
         dconv1d_out = rearrange(dconv1d_out, "d (b l) -> b d l", b=x.shape[0], l=x.shape[-1])
         # The kernel supports passing in a pre-allocated dx (e.g., in case we want to fuse the
         # backward of conv1d with the backward of chunk).
-        dx, dconv1d_weight, dconv1d_bias = causal_conv1d_cuda.causal_conv1d_bwd(
-            x, conv1d_weight, conv1d_bias, dconv1d_out, None, dx, True
+        dx, dconv1d_weight, dconv1d_bias, *_ = causal_conv1d_cuda.causal_conv1d_bwd(
+            x, conv1d_weight, conv1d_bias, dconv1d_out, None, None, None, dx, False, True
         )
         dconv1d_bias = dconv1d_bias if conv1d_bias is not None else None
         dconv1d_weight = rearrange(dconv1d_weight, "d w -> d 1 w")
@@ -374,11 +381,12 @@ def mamba_inner_ref(
     A, B=None, C=None, D=None, delta_bias=None, B_proj_bias=None,
     C_proj_bias=None, delta_softplus=True
 ):
+    assert causal_conv1d_fn is not None, "causal_conv1d_fn is not available. Please install causal-conv1d."
     L = xz.shape[-1]
     delta_rank = delta_proj_weight.shape[1]
     d_state = A.shape[-1] * (1 if not A.is_complex() else 2)
     x, z = xz.chunk(2, dim=1)
-    x = causal_conv1d_fn(x, rearrange(conv1d_weight, "d 1 w -> d w"), conv1d_bias, "silu")
+    x = causal_conv1d_fn(x, rearrange(conv1d_weight, "d 1 w -> d w"), conv1d_bias, activation="silu")
     # We're being very careful here about the layout, to avoid extra transposes.
     # We want delta to have d as the slowest moving dimension
     # and L as the fastest moving dimension, since those are what the ssm_scan kernel expects.

setup.py

@@ -271,6 +271,6 @@ def run(self):
         "einops",
         "triton",
         "transformers",
-        "causal_conv1d>=1.1.0,<1.2.0",
+        # "causal_conv1d>=1.2.0",
     ],
 )

tests/ops/triton/test_selective_state_update.py

@@ -19,7 +19,7 @@
 # @pytest.mark.parametrize("dstate", [16])
 @pytest.mark.parametrize("dim", [2048, 2048 + 16, 4096])
 # @pytest.mark.parametrize("dim", [2048])
-def test_causal_conv1d_update(dim, dstate, has_z, itype):
+def test_selective_state_update(dim, dstate, has_z, itype):
     device = "cuda"
     rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (5e-3, 1e-2)
     if itype == torch.bfloat16: