Add ApplyRoPE and RMSNorm kernels written in OpenAI Triton

NOTICE

@@ -49,4 +49,232 @@ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+SOFTWARE.
+
+------------- LICENSE FOR stanford_alpaca code  --------------
+
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+   
+------------- LICENSE FOR PanQiWei AutoGPTQ code  --------------
+
+MIT License
+
+Copyright (c) 2023 潘其威(William)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -16,11 +16,11 @@ inference: false
 <br>
 
 <p align="center">
-        🤗 <a href="https://huggingface.co/Qwen">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/qwen">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://arxiv.org/abs/2309.16609">Paper</a>&nbsp&nbsp ｜ &nbsp&nbsp🖥️ <a href="https://modelscope.cn/studios/qwen/Qwen-7B-Chat-Demo/summary">Demo</a>
+        🤗 <a href="https://huggingface.co/Qwen">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/qwen">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://arxiv.org/abs/2309.16609">Paper</a> &nbsp&nbsp ｜ &nbsp&nbsp🖥️ <a href="https://modelscope.cn/studios/qwen/Qwen-7B-Chat-Demo/summary">Demo</a>
 <br>
-<a href="https://github.com/QwenLM/Qwen/blob/main/assets/wechat.png">WeChat (微信)</a>&nbsp&nbsp ｜ &nbsp&nbsp DingTalk (钉钉) &nbsp&nbsp | &nbsp&nbsp<a href="https://discord.gg/z3GAxXZ9Ce">Discord</a>&nbsp&nbsp
+<a href="assets/wechat.png">WeChat (微信)</a>&nbsp&nbsp | &nbsp&nbsp<a href="https://discord.gg/z3GAxXZ9Ce">Discord</a>&nbsp&nbsp ｜  &nbsp&nbsp<a href="https://dashscope.aliyun.com">API</a> 
 </p>
-<br><br>
+<br>
 
 ## 介绍（Introduction）
 
@@ -597,9 +597,9 @@ If you find our work helpful, feel free to give us a cite.
 
 ## 使用协议（License Agreement）
 
-我们的代码和模型权重对学术研究完全开放，并支持商用。请查看[LICENSE](https://github.com/QwenLM/Qwen/blob/main/LICENSE)了解具体的开源协议细节。如需商用，请填写[问卷](https://dashscope.console.aliyun.com/openModelApply/qianwen)申请。
+我们的代码和模型权重对学术研究完全开放，并支持商用。请查看[LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)了解具体的开源协议细节。如需商用，请填写[问卷](https://dashscope.console.aliyun.com/openModelApply/qianwen)申请。
 
-Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check [LICENSE](https://github.com/QwenLM/Qwen/blob/main/LICENSE) for more details about the license. If you have requirements for commercial use, please fill out the [form](https://dashscope.console.aliyun.com/openModelApply/qianwen) to apply.
+Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check [LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) for more details about the license. If you have requirements for commercial use, please fill out the [form](https://dashscope.console.aliyun.com/openModelApply/qianwen) to apply.
 <br>
 
 

config.json

@@ -16,7 +16,7 @@
   "initializer_range": 0.02,
   "kv_channels": 128,
   "layer_norm_epsilon": 1e-06,
-  "max_position_embeddings": 8192,
+  "max_position_embeddings": 32768,
   "model_type": "qwen",
   "no_bias": true,
   "num_attention_heads": 32,
@@ -44,6 +44,7 @@
   "use_cache": true,
   "use_dynamic_ntk": true,
   "use_flash_attn": "auto",
+  "use_triton": "auto",
   "use_logn_attn": true,
   "vocab_size": 151936
 }

configuration_qwen.py

@@ -32,6 +32,7 @@ class QWenConfig(PretrainedConfig):
         use_dynamic_ntk=True,
         use_logn_attn=True,
         use_flash_attn="auto",
+        use_triton="auto",
         intermediate_size=22016,
         no_bias=True,
         tie_word_embeddings=False,
@@ -61,6 +62,7 @@ class QWenConfig(PretrainedConfig):
         self.use_dynamic_ntk = use_dynamic_ntk
         self.use_logn_attn = use_logn_attn
         self.use_flash_attn = use_flash_attn
+        self.use_triton = use_triton
         self.no_bias = no_bias
         self.use_cache_quantization = use_cache_quantization
         self.use_cache_kernel = use_cache_kernel

generation_config.json

@@ -1,12 +1,12 @@
 {
-  "chat_format": "chatml",
-  "eos_token_id": 151643,
-  "pad_token_id": 151643,
-  "max_window_size": 6144,
-  "max_new_tokens": 512,
-  "do_sample": true,
-  "top_k": 0,
-  "top_p": 0.8,
-  "repetition_penalty": 1.1,
-  "transformers_version": "4.31.0"
-}
+    "chat_format": "chatml",
+    "eos_token_id": 151643,
+    "pad_token_id": 151643,
+    "max_window_size": 24000,
+    "max_new_tokens": 512,
+    "do_sample": true,
+    "top_k": 0,
+    "top_p": 0.8,
+    "repetition_penalty": 1.1,
+    "transformers_version": "4.31.0"
+  }

modeling_qwen.py

@@ -13,7 +13,6 @@ import torch
 import torch.nn.functional as F
 import torch.utils.checkpoint
 import warnings
-from torch.cuda.amp import autocast
 
 from torch.nn import CrossEntropyLoss
 from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
@@ -36,7 +35,7 @@ except ImportError:
 from torch import nn
 
 SUPPORT_CUDA = torch.cuda.is_available()
-SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
+SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 8
 SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
 SUPPORT_TORCH2 = hasattr(torch, '__version__') and int(torch.__version__.split(".")[0]) >= 2
 
@@ -77,11 +76,14 @@ We detect you have activated flash attention support, but running model computat
 """
 
 apply_rotary_emb_func = None
+apply_rotary_emb_func_triton = None
 rms_norm = None
+rms_norm_triton = None
 flash_attn_unpadded_func = None
+flash_attn_func = None
 
 def _import_flash_attn():
-    global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
+    global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func, flash_attn_func
     try:
         from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
         apply_rotary_emb_func = __apply_rotary_emb_func
@@ -102,20 +104,42 @@ def _import_flash_attn():
 
     try:
         import flash_attn
+        _flash_attn_func = None
         if not hasattr(flash_attn, '__version__'):
             from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
         else:
             if int(flash_attn.__version__.split(".")[0]) >= 2:
+                if int(flash_attn.__version__.split(".")[1]) >= 1:
+                    from flash_attn.flash_attn_interface import flash_attn_func as _flash_attn_func
                 from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
             else:
                 from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
         flash_attn_unpadded_func = __flash_attn_unpadded_func
+        flash_attn_func = _flash_attn_func
     except ImportError:
         logger.warn(
             "Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
             "https://github.com/Dao-AILab/flash-attention"
         )
 
+def _import_triton():
+    global apply_rotary_emb_func_triton, rms_norm_triton
+    try:
+        from .triton_kernels import apply_rotary_emb as __apply_rotary_emb, rms_norm as __rms_norm
+        if apply_rotary_emb_func is not None:
+            logger.warn(
+                "Using Triton rotary kernel instead of flash_attn for inference."
+            )
+        apply_rotary_emb_func_triton = __apply_rotary_emb
+        if rms_norm is not None:
+            logger.warn(
+                "Using Triton rms_norm kernel instead of flash_attn for inference."
+            )
+        rms_norm_triton = __rms_norm
+    except ImportError:
+        logger.warn("Warning: Failed to import Triton kernels.")
+        return
+
 def quantize_cache_v(fdata, bits, qmax, qmin):
     # b, s, head, h-dim->b, head, s, h-dim
     qtype = torch.uint8
@@ -182,6 +206,11 @@ class FlashSelfAttention(torch.nn.Module):
         seqlen_k = k.shape[1]
         seqlen_out = seqlen_q
 
+        if flash_attn_func is not None and batch_size == 1:
+            dropout_p = self.dropout_p if self.training else 0
+            output = flash_attn_func(q, k, v, dropout_p, softmax_scale=self.softmax_scale, causal=self.causal)
+            return output
+
         q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
         cu_seqlens_q = torch.arange(
             0,
@@ -311,7 +340,7 @@ class QWenAttention(nn.Module):
                     warnings.warn("Failed to import KV cache kernels.")
                     self.cache_kernels = None
 
-    def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
+    def _attn(self, query, key, value, causal_mask=None, attention_mask=None, head_mask=None):
         device = query.device
         if self.use_cache_quantization:
             qk, qk_scale, qk_zero = key
@@ -336,26 +365,13 @@ class QWenAttention(nn.Module):
                 size_temp = value[0].size(-1)
             else:
                 size_temp = value.size(-1)
-            attn_weights = attn_weights / torch.full(
-                [],
-                size_temp ** 0.5,
-                dtype=attn_weights.dtype,
-                device=attn_weights.device,
-            )
-        if self.use_cache_quantization:
-            query_length, key_length = query.size(-2), key[0].size(-2)
-        else:
-            query_length, key_length = query.size(-2), key.size(-2)
-        causal_mask = registered_causal_mask[
-            :, :, key_length - query_length : key_length, :key_length
-        ]
+            attn_weights = attn_weights / (size_temp ** 0.5)
+
         mask_value = torch.finfo(attn_weights.dtype).min
-        mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
-            attn_weights.device
-        )
-        attn_weights = torch.where(
-            causal_mask, attn_weights.to(attn_weights.dtype), mask_value
-        )
+        if causal_mask is not None:
+            attn_weights = torch.where(
+                causal_mask, attn_weights.to(attn_weights.dtype), mask_value
+            )
 
         if attention_mask is not None:
             attn_weights = attn_weights + attention_mask
@@ -395,62 +411,6 @@ class QWenAttention(nn.Module):
 
         return attn_output, attn_weights
 
-    def _upcast_and_reordered_attn(
-        self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None
-    ):
-        bsz, num_heads, q_seq_len, dk = query.size()
-        _, _, k_seq_len, _ = key.size()
-
-        attn_weights = torch.empty(
-            bsz * num_heads,
-            q_seq_len,
-            k_seq_len,
-            dtype=torch.float32,
-            device=query.device,
-        )
-
-        scale_factor = 1.0
-        if self.scale_attn_weights:
-            scale_factor /= float(value.size(-1)) ** 0.5
-
-        with autocast(enabled=False):
-            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
-                -1, dk, k_seq_len
-            )
-            attn_weights = torch.baddbmm(
-                attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
-            )
-            attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
-
-        query_length, key_length = query.size(-2), key.size(-2)
-        causal_mask = registered_causal_mask[
-            :, :, key_length - query_length : key_length, :key_length
-        ]
-        mask_value = torch.finfo(attn_weights.dtype).min
-        mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(
-            attn_weights.device
-        )
-        attn_weights = torch.where(causal_mask, attn_weights, mask_value)
-
-        if attention_mask is not None:
-            attn_weights = attn_weights + attention_mask
-
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
-
-        if attn_weights.dtype != torch.float32:
-            raise RuntimeError(
-                "Error with upcasting, attn_weights does not have dtype torch.float32"
-            )
-        attn_weights = attn_weights.type(value.dtype)
-        attn_weights = self.attn_dropout(attn_weights)
-
-        if head_mask is not None:
-            attn_weights = attn_weights * head_mask
-
-        attn_output = torch.matmul(attn_weights, value)
-
-        return attn_output, attn_weights
-
     def _split_heads(self, tensor, num_heads, attn_head_size):
         new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
         tensor = tensor.view(new_shape)
@@ -465,7 +425,6 @@ class QWenAttention(nn.Module):
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
         rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
-        registered_causal_mask: Optional[torch.Tensor] = None,
         layer_past: Optional[Tuple[torch.Tensor]] = None,
         attention_mask: Optional[torch.FloatTensor] = None,
         head_mask: Optional[torch.FloatTensor] = None,
@@ -539,7 +498,8 @@ class QWenAttention(nn.Module):
         else:
             present = None
 
-        if self.use_logn_attn and not self.training:
+        key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
+        if key_size > self.seq_length and self.use_logn_attn and not self.training:
             if self.use_cache_quantization:
                 seq_start = key[0].size(2) - query.size(1)
                 seq_end = key[0].size(2)
@@ -558,12 +518,19 @@ class QWenAttention(nn.Module):
             q, k, v = query, key, value
             attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
         else:
+            key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
+            if query.size(1) == key_size:
+                causal_mask = torch.tril(
+                    torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
+                ).view(1, 1, key_size, key_size)
+            else:
+                causal_mask = None
             query = query.permute(0, 2, 1, 3)
             if not self.use_cache_quantization:
                 key = key.permute(0, 2, 1, 3)
                 value = value.permute(0, 2, 1, 3)
             if (
-                registered_causal_mask is None
+                causal_mask is None
                 and self.use_flash_attn
                 and flash_attn_unpadded_func is not None
                 and not self.is_fp32
@@ -572,13 +539,12 @@ class QWenAttention(nn.Module):
                 raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
 
             if not self.use_cache_quantization and SUPPORT_TORCH2:
-                causal_mask = registered_causal_mask[
-                    :, :, key.size(-2) - query.size(-2): key.size(-2), :key.size(-2)
-                ]
                 if attention_mask is not None:
                     attention_mask = attention_mask.expand(
                         -1, -1, causal_mask.size(2), -1
-                    ).masked_fill(~causal_mask, torch.finfo(query.dtype).min)
+                    )
+                    if causal_mask is not None:
+                        attention_mask.masked_fill_(~causal_mask, torch.finfo(query.dtype).min)
                 else:
                     attention_mask = causal_mask
                 attn_output = F.scaled_dot_product_attention(
@@ -587,7 +553,7 @@ class QWenAttention(nn.Module):
                 attn_weight = None
             else:
                 attn_output, attn_weight = self._attn(
-                    query, key, value, registered_causal_mask, attention_mask, head_mask
+                    query, key, value, causal_mask, attention_mask, head_mask
                 )
         context_layer = self._merge_heads(
             attn_output, self.num_heads, self.head_dim
@@ -603,6 +569,8 @@ class QWenAttention(nn.Module):
                 and not self.is_fp32
             ):
                 raise ValueError("Cannot output attentions while using flash-attn")
+            elif not self.use_cache_quantization and SUPPORT_TORCH2:
+                raise ValueError("Cannot output attentions while using scaled_dot_product_attention")
             else:
                 outputs += (attn_weight,)
 
@@ -628,6 +596,7 @@ class QWenMLP(nn.Module):
         output = self.c_proj(intermediate_parallel)
         return output
 
+
 class QWenBlock(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -650,7 +619,6 @@ class QWenBlock(nn.Module):
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
         rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
-        registered_causal_mask: Optional[torch.Tensor] = None,
         layer_past: Optional[Tuple[torch.Tensor]] = None,
         attention_mask: Optional[torch.FloatTensor] = None,
         head_mask: Optional[torch.FloatTensor] = None,
@@ -664,7 +632,6 @@ class QWenBlock(nn.Module):
         attn_outputs = self.attn(
             layernorm_output,
             rotary_pos_emb_list,
-            registered_causal_mask=registered_causal_mask,
             layer_past=layer_past,
             attention_mask=attention_mask,
             head_mask=head_mask,
@@ -698,6 +665,7 @@ class QWenPreTrainedModel(PreTrainedModel):
     is_parallelizable = False
     supports_gradient_checkpointing = True
     _no_split_modules = ["QWenBlock"]
+    _skip_keys_device_placement = "past_key_values"
 
     def __init__(self, *inputs, **kwargs):
         super().__init__(*inputs, **kwargs)
@@ -764,21 +732,6 @@ class QWenModel(QWenPreTrainedModel):
 
         self.use_flash_attn = config.use_flash_attn
         self.is_fp32 = not (config.bf16 or config.fp16)
-        if (
-            self.use_flash_attn
-            and flash_attn_unpadded_func is not None
-            and not self.is_fp32
-        ):
-            self.registered_causal_mask = None
-        else:
-            max_positions = config.max_position_embeddings
-            self.register_buffer(
-                "registered_causal_mask",
-                torch.tril(
-                    torch.ones((max_positions, max_positions), dtype=torch.bool)
-                ).view(1, 1, max_positions, max_positions),
-                persistent=False,
-            )
 
         self.h = nn.ModuleList(
             [
@@ -950,7 +903,6 @@ class QWenModel(QWenPreTrainedModel):
                     create_custom_forward(block),
                     hidden_states,
                     rotary_pos_emb_list,
-                    self.registered_causal_mask,
                     None,
                     attention_mask,
                     head_mask[i],
@@ -962,7 +914,6 @@ class QWenModel(QWenPreTrainedModel):
                     hidden_states,
                     layer_past=layer_past,
                     rotary_pos_emb_list=rotary_pos_emb_list,
-                    registered_causal_mask=self.registered_causal_mask,
                     attention_mask=attention_mask,
                     head_mask=head_mask[i],
                     encoder_hidden_states=encoder_hidden_states,
@@ -1006,11 +957,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         assert (
             config.bf16 + config.fp16 + config.fp32 <= 1
         ), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
-        logger.warn(
-            "Warning: please make sure that you are using the latest codes and checkpoints, "
-            "especially if you used Qwen-7B before 09.25.2023."
-            "请使用最新模型和代码，尤其如果你在9月25日前已经开始使用Qwen-7B，千万注意不要使用错误代码和模型。"
-        )
 
         autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
 
@@ -1052,6 +998,12 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         if config.use_flash_attn:
             _import_flash_attn()
 
+        if config.use_triton == "auto":
+            logger.warn("Try importing Triton kernels for faster inference...")
+            config.use_triton = SUPPORT_TORCH2
+        if config.use_triton:
+            _import_triton()
+
         self.transformer = QWenModel(config)
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
 
@@ -1063,7 +1015,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             self.lm_head.half()
         self.post_init()
 
-
     def get_output_embeddings(self):
         return self.lm_head
 
@@ -1073,22 +1024,13 @@ class QWenLMHeadModel(QWenPreTrainedModel):
     def prepare_inputs_for_generation(
         self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
     ):
-        token_type_ids = kwargs.get("token_type_ids", None)
         if past_key_values:
             input_ids = input_ids[:, -1].unsqueeze(-1)
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
-
-        attention_mask = kwargs.get("attention_mask", None)
-        position_ids = kwargs.get("position_ids", None)
 
-        if attention_mask is not None and position_ids is None:
-            position_ids = attention_mask.long().cumsum(-1) - 1
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if past_key_values:
-                position_ids = position_ids[:, -1].unsqueeze(-1)
+        if input_ids.size(0) == 1:
+            attention_mask = None
         else:
-            position_ids = None
+            attention_mask = kwargs.get("attention_mask", None)
 
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}
@@ -1099,9 +1041,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             {
                 "past_key_values": past_key_values,
                 "use_cache": kwargs.get("use_cache"),
-                "position_ids": position_ids,
                 "attention_mask": attention_mask,
-                "token_type_ids": token_type_ids,
             }
         )
         return model_inputs
@@ -1372,8 +1312,7 @@ class RotaryEmbedding(torch.nn.Module):
         self._ntk_alpha_cached = 1.0
         self._ntk_alpha_cached_list = [1.0]
 
-    def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
-        seqlen = max_seq_len + offset
+    def update_rotary_pos_emb_cache(self, seqlen, ntk_alpha=1.0):
         if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
             base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
             self.inv_freq = 1.0 / (
@@ -1396,10 +1335,10 @@ class RotaryEmbedding(torch.nn.Module):
             cos, sin = emb.cos(), emb.sin()
             self._rotary_pos_emb_cache = [cos, sin]
 
-    def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
-        self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
+    def forward(self, max_seq_len, ntk_alpha=1.0):
+        self.update_rotary_pos_emb_cache(max_seq_len, ntk_alpha)
         cos, sin = self._rotary_pos_emb_cache
-        return [cos[:, offset : offset + max_seq_len], sin[:, offset : offset + max_seq_len]]
+        return [cos[:, :max_seq_len], sin[:, :max_seq_len]]
 
 
 def _rotate_half(x):
@@ -1411,21 +1350,30 @@ def _rotate_half(x):
 
 
 def apply_rotary_pos_emb(t, freqs):
+    """ Apply rotary embedding to the first rotary_dim of the iput
+
+    Arguments:
+      t (tensor(batch_size, seq_len, n_head, head_dim)):
+        the input embedding/hidden states
+      freqs (list[tensor(1, seq_len, 1, rotary_dim), tensor(1, seq_len, 1, rotary_dim)]):
+        the cached cos/sin position embeddings 
+    """
+    rot_dim = freqs[0].shape[-1]
     cos, sin = freqs
-    if apply_rotary_emb_func is not None and t.is_cuda:
-        t_ = t.float()
-        cos = cos.squeeze(0).squeeze(1)[:, : cos.shape[-1] // 2]
-        sin = sin.squeeze(0).squeeze(1)[:, : sin.shape[-1] // 2]
-        output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
-        return output
+    t_float = t.float()
+    if apply_rotary_emb_func_triton is not None and t.is_cuda and (not t.requires_grad):
+        return apply_rotary_emb_func_triton(t, cos, sin)
+    elif apply_rotary_emb_func is not None and t.is_cuda:
+        # apply_rotary_emb in flash_attn requires cos/sin to be of 
+        # shape (seqlen, rotary_dim / 2) and apply rotary embedding 
+        # to the first rotary_dim of the input
+        cos = cos.squeeze(0).squeeze(1)[:, : rot_dim // 2]
+        sin = sin.squeeze(0).squeeze(1)[:, : rot_dim // 2]
+        return apply_rotary_emb_func(t_float, cos, sin).type_as(t)
     else:
-        rot_dim = freqs[0].shape[-1]
-        cos, sin = freqs
-        t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
-        t_ = t_.float()
-        t_pass_ = t_pass_.float()
-        t_ = (t_ * cos) + (_rotate_half(t_) * sin)
-        return torch.cat((t_, t_pass_), dim=-1).type_as(t)
+        t_rot, t_pass = t_float[..., :rot_dim], t_float[..., rot_dim:]
+        t_rot = (t_rot * cos) + (_rotate_half(t_rot) * sin)
+        return torch.cat((t_rot, t_pass), dim=-1).type_as(t)
 
 
 class RMSNorm(torch.nn.Module):
@@ -1438,7 +1386,9 @@ class RMSNorm(torch.nn.Module):
         return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
 
     def forward(self, x):
-        if rms_norm is not None and x.is_cuda:
+        if rms_norm_triton is not None and x.is_cuda and (not x.requires_grad):
+            return rms_norm_triton(x, self.weight, self.eps)
+        elif rms_norm is not None and x.is_cuda:
             return rms_norm(x, self.weight, self.eps)
         else:
             output = self._norm(x.float()).type_as(x)

tokenizer_config.json

@@ -1,5 +1,5 @@
 {
-  "model_max_length": 8192,
+  "model_max_length": 32768,
   "tokenizer_class": "QWenTokenizer",
   "auto_map": {
     "AutoTokenizer": [

triton_kernels.py

@@ -0,0 +1,115 @@
+from typing import Any, Callable, Dict, Hashable, Tuple
+
+import torch
+import triton
+import triton.language as tl
+from triton.compiler import CompiledKernel
+from triton.runtime import JITFunction
+
+try:
+    import triton.language.math as tlmath  # Triton 2.1
+except ImportError:
+    import triton.language.libdevice as tlmath  # Triton 2.0
+
+
+class TritonKernel:
+    def __init__(
+        self,
+        kernel_fn: JITFunction,
+        grid_fn: Callable[[Tuple[Any, ...]], Tuple[int, int, int]],
+    ) -> None:
+        self.kernel_fn_ = kernel_fn
+        self.grid_fn_ = grid_fn
+        self.kernel_cache_: Dict[Hashable, CompiledKernel] = {}
+
+    def run(self, *args, **kwargs):
+        # Set current device
+        input_device = args[0].device
+        prev_dev_idx, cur_dev_idx = -1, torch.cuda.current_device()
+        if input_device.index != cur_dev_idx:
+            prev_dev_idx = cur_dev_idx
+            torch.cuda.set_device(input_device.index)
+
+        # Compute grid
+        grid = self.grid_fn_(args)
+
+        # Use cached kernel if possible
+        kernel_key = (input_device,) + tuple(kwargs.items())
+        if kernel_key in self.kernel_cache_:
+            kernel = self.kernel_cache_[kernel_key]
+            kernel[grid](*args)
+        else:
+            # Compile and store new kernel
+            kernel = self.kernel_fn_[grid](*args, **kwargs)
+            self.kernel_cache_[kernel_key] = kernel
+
+        # Restore previous device
+        torch.cuda.set_device(prev_dev_idx)
+
+
+@triton.jit
+def _apply_rope_fwd_kernel(X, Cos, Sin, Y, HEAD_DIM: tl.constexpr):
+    batch_idx, tok_idx, head_idx = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    seq_len, num_heads = tl.num_programs(1), tl.num_programs(2)
+    block_idx = tl.arange(0, HEAD_DIM)
+    x_base_idx = ((batch_idx * seq_len + tok_idx) * num_heads * 3 + head_idx) * HEAD_DIM
+    x = tl.load(X + x_base_idx + block_idx)
+    freq_idx = tok_idx * HEAD_DIM + block_idx
+    cos = tl.load(Cos + freq_idx)
+    rot_idx = (HEAD_DIM // 2 + block_idx) % HEAD_DIM
+    x_rot = tl.load(X + x_base_idx + rot_idx)
+    x_rot = tl.where(block_idx >= HEAD_DIM // 2, x_rot, -x_rot)
+    sin = tl.load(Sin + freq_idx)
+    y_idx = (
+        (batch_idx * seq_len + tok_idx) * num_heads + head_idx
+    ) * HEAD_DIM + block_idx
+    y = x * cos + x_rot * sin
+    tl.store(Y + y_idx, y.to(Y.dtype.element_ty))
+
+
+apply_rope_fwd_kernel = TritonKernel(
+    _apply_rope_fwd_kernel, lambda args: tuple(args[0].shape[:3])
+)
+
+
+def apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor):
+    y = torch.empty(x.shape, dtype=x.dtype, device=x.device)
+    apply_rope_fwd_kernel.run(x, cos, sin, y, HEAD_DIM=x.size(-1))
+    return y
+
+
+@triton.jit
+def _rms_norm_fwd_kernel(X, W, Y, eps, hidden_dim, BLOCK_SIZE: tl.constexpr):
+    tok_idx = tl.program_id(0)
+
+    mean_sq = tl.zeros([BLOCK_SIZE], tl.float32)
+    for offset in range(0, hidden_dim, BLOCK_SIZE):
+        dim_idx = offset + tl.arange(0, BLOCK_SIZE)
+        x = tl.load(
+            X + tok_idx * hidden_dim + dim_idx, mask=dim_idx < hidden_dim, other=0
+        ).to(tl.float32)
+        mean_sq += x * x / hidden_dim
+    rrms = tlmath.rsqrt(tl.sum(mean_sq, 0) + eps)
+
+    for offset in range(0, hidden_dim, BLOCK_SIZE):
+        dim_idx = offset + tl.arange(0, BLOCK_SIZE)
+        dim_mask = dim_idx < hidden_dim
+        hidden_idx = tok_idx * hidden_dim + dim_idx
+        x = tl.load(X + hidden_idx, mask=dim_mask, other=0)
+        w = tl.load(W + dim_idx, mask=dim_mask, other=0)
+        y = x * rrms * w
+        tl.store(Y + hidden_idx, y.to(Y.dtype.element_ty), mask=dim_mask)
+
+
+rms_norm_fwd_kernel = TritonKernel(
+    _rms_norm_fwd_kernel, lambda args: (args[0].shape[:-1].numel(), 1, 1)
+)
+
+
+def rms_norm(x: torch.Tensor, weight: torch.Tensor, eps: float):
+    y = torch.empty_like(x)
+    hidden_dim = x.size(-1)
+    rms_norm_fwd_kernel.run(
+        x, weight, y, eps, hidden_dim, BLOCK_SIZE=triton.next_power_of_2(hidden_dim)
+    )
+    return y