[v1] support GDN Ulysses cp (#10727)

Co-authored-by: cxy-thinkbook <xuanyuchen@seu.edu.cn>

data/v1_multimodal_demo.jsonl

@@ -22,4 +22,3 @@
 {"messages": [{"role": "user", "content": [{"type": "image_url", "value": "data/mllm_demo_data/1.jpg"}, {"type": "text", "value": "他们是谁？"}]}, {"role": "assistant", "content": [{"type": "text", "value": "他们是拜仁慕尼黑的凯恩和格雷茨卡。"}]}, {"role": "user", "content": [{"type": "text", "value": "他们在做什么？"}, {"type": "image_url", "value": "data/mllm_demo_data/1.jpg"}]}, {"role": "assistant", "content": [{"type": "text", "value": "他们在足球场上庆祝。"}]}]}
 {"messages": [{"role": "user", "content": [{"type": "image_url", "value": "data/mllm_demo_data/2.jpg"}, {"type": "text", "value": "他是谁？"}]}, {"role": "assistant", "content": [{"type": "text", "value": "他是来自拜仁慕尼黑的托马斯·穆勒。"}]}, {"role": "user", "content": [{"type": "text", "value": "他为什么在地上？"}]}, {"role": "assistant", "content": [{"type": "text", "value": "因为他正在双膝跪地滑行庆祝。"}]}]}
 {"messages": [{"role": "user", "content": [{"type": "image_url", "value": "data/mllm_demo_data/3.jpg"}, {"type": "text", "value": "请描述这张图片"}]}, {"role": "assistant", "content": [{"type": "text", "value": "中国宇航员桂海潮正在讲话。"}]}, {"role": "user", "content": [{"type": "text", "value": "他取得过哪些成就？"}]}, {"role": "assistant", "content": [{"type": "text", "value": "他于2022年6月被任命为神舟十六号任务的有效载荷专家，从而成为2023年5月30日进入太空的首位平民宇航员。他负责在轨操作空间科学实验有效载荷。"}]}]}
-

data/v1_multimodal_demo.yaml

@@ -1,4 +1,3 @@
 multimodal_demo:
   path: data/v1_multimodal_demo.jsonl
   source: local
-

examples/v1/train_full/train_multimodal.yaml

@@ -24,4 +24,3 @@ max_steps: 5
 ### sample
 sample_backend: hf
 max_new_tokens: 128
-

src/llamafactory/v1/core/base_trainer.py

@@ -147,11 +147,6 @@ class BaseTrainer:
         self.state.epoch = self._resume_epoch
 
         if self.args.cp_size > 1:
-            # qwen3.5 is not supported because of the different attention implementation, which will be supported in the future.
-            if model.config.model_type == "qwen3_5":
-                raise RuntimeError(
-                    "Sequence parallel is not supported for qwen3.5 model due to its different attention implementation, which will be supported in the future."
-                )
             from ..plugins.model_plugins.parallelization.sequence_parallel import SequenceParallelModelPlugin
 
             if model.config._attn_implementation != "flash_attention_2":

src/llamafactory/v1/core/rendering/rendering.py

@@ -23,7 +23,6 @@ Note: ``position_ids`` are assigned by ``process_samples`` (1-based); multimodal
 ids are expected to be recomputed by the model/trainer.
 """
 
-
 import json
 
 import numpy as np

src/llamafactory/v1/plugins/data_plugins/converter.py

@@ -295,4 +295,3 @@ def pair_converter(raw_sample: PairSample) -> DPOSample:
             logger.warning_rank0(f"Invalid tools format: {str(tools)}")
 
     return sample
-

src/llamafactory/v1/plugins/model_plugins/parallelization/gdn_attention.py

@@ -0,0 +1,250 @@
+# Copyright 2025 the LlamaFactory team.
+#
+# 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.
+
+
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+
+from .seq_comm import SeqAllToAll4D
+from .ulysses import (
+    get_ulysses_sequence_parallel_group,
+    get_ulysses_sequence_parallel_world_size,
+)
+
+
+def is_gdn_layer(layer) -> bool:
+    """Return True if the module is a GDN (linear attention) layer or a DecoderLayer containing one."""
+    if hasattr(layer, "layer_type") and layer.layer_type == "linear_attention":
+        return True
+    if hasattr(layer, "block_type") and layer.block_type == "linear_attention":
+        return True
+    return False
+
+
+def _get_gdn_module(module):
+    """Return the actual GDN module from either a GDN layer or a DecoderLayer."""
+    if hasattr(module, "in_proj_qkv"):
+        return module
+    if hasattr(module, "linear_attn"):
+        return module.linear_attn
+    raise AttributeError(
+        f"Cannot find GDN module on {type(module).__name__}. "
+        f"Expected either a GDN layer with in_proj_qkv or a DecoderLayer with linear_attn."
+    )
+
+
+def get_parameter_local_cp(param, dim, cp_group, split_sections=None):
+    """Slice a parameter for the current CP rank.
+
+    If split_sections is given, first split along dim into sub-groups,
+    slice each sub-group independently for CP, then concatenate back.
+    This ensures each CP rank gets a proportional slice of each sub-group.
+    """
+    cp_size = dist.get_world_size(group=cp_group)
+    if cp_size == 1:
+        return param
+    cp_rank = dist.get_rank(group=cp_group)
+
+    if split_sections is not None:
+        inputs = torch.split(param, split_sections, dim=dim)
+        outputs = []
+        for p in inputs:
+            p = get_parameter_local_cp(p, dim, cp_group)
+            outputs.append(p)
+        return torch.cat(outputs, dim=dim)
+
+    slices = [slice(None)] * param.dim()
+    dim_size = param.size(dim=dim)
+    slices[dim] = slice(cp_rank * dim_size // cp_size, (cp_rank + 1) * dim_size // cp_size)
+    return param[slices]
+
+
+def gdn_forward_with_cp(self, hidden_states, attention_mask=None, **kwargs):
+    """GDN forward with Context Parallel support.
+
+    Uses SeqAllToAll4D (same as UlyssesAttention) for all all_to_all operations.
+    Each component (Q/K/V/z/b/a) is independently reshaped to 4D and all_to_all'd
+    with scatter heads / gather seq, avoiding the bug where uniform hidden-split on
+    merged qkv gives rank-0 [Q+K] and rank-1 [V].
+
+    Falls back to self.original_forward when cp_size <= 1.
+    """
+    cp_size = get_ulysses_sequence_parallel_world_size()
+    if cp_size <= 1:
+        return self.original_forward(hidden_states, attention_mask=attention_mask, **kwargs)
+
+    cp_group = get_ulysses_sequence_parallel_group()
+
+    if attention_mask is not None:
+        try:
+            from transformers.models.qwen3_5.modeling_qwen3_5 import apply_mask_to_padding_states
+
+            hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
+        except ImportError:
+            pass
+
+    batch_size, seq_len, _ = hidden_states.shape
+    full_seq_len = seq_len * cp_size
+
+    # Extract position_ids and derive cu_seqlens for pack support
+    position_ids = kwargs.get("position_ids", None)
+    cu_seqlens = None
+    if position_ids is not None and batch_size == 1:
+        global_position_ids = [torch.empty_like(position_ids) for _ in range(cp_size)]
+        dist.all_gather(global_position_ids, position_ids, group=cp_group)
+        global_position_ids = torch.cat(global_position_ids, dim=-1).contiguous()
+        try:
+            from transformers.modeling_flash_attention_utils import prepare_fa_kwargs_from_position_ids
+            cu_seqlens = prepare_fa_kwargs_from_position_ids(global_position_ids)[0][0]
+        except ImportError:
+            cu_seqlens = None
+
+    # Input projections in CP layout: [B, seq/cp, hidden]
+    qkv = self.in_proj_qkv(hidden_states)  # [B, seq/cp, key_dim*2 + value_dim]
+    z = self.in_proj_z(hidden_states)  # [B, seq/cp, value_dim]
+    b = self.in_proj_b(hidden_states)  # [B, seq/cp, num_v_heads]
+    a = self.in_proj_a(hidden_states)  # [B, seq/cp, num_v_heads]
+
+    # Split qkv into Q, K, V before all_to_all so each sub-group gets
+    # proportional head distribution across ranks.
+    q_proj, k_proj, v_proj = torch.split(qkv, [self.key_dim, self.key_dim, self.value_dim], dim=-1)
+
+    # CP->HP all_to_all for each component: scatter heads (dim=2), gather seq (dim=1)
+    # [B, S/cp, heads, head_dim] -> [B, S, heads/cp, head_dim]
+    q_proj = q_proj.reshape(batch_size, seq_len, self.num_k_heads, self.head_k_dim)
+    q_proj = SeqAllToAll4D.apply(cp_group, q_proj, 2, 1)
+
+    k_proj = k_proj.reshape(batch_size, seq_len, self.num_k_heads, self.head_k_dim)
+    k_proj = SeqAllToAll4D.apply(cp_group, k_proj, 2, 1)
+
+    v_proj = v_proj.reshape(batch_size, seq_len, self.num_v_heads, self.head_v_dim)
+    v_proj = SeqAllToAll4D.apply(cp_group, v_proj, 2, 1)
+
+    z = z.reshape(batch_size, seq_len, self.num_v_heads, self.head_v_dim)
+    z = SeqAllToAll4D.apply(cp_group, z, 2, 1)
+
+    b = b.reshape(batch_size, seq_len, self.num_v_heads, 1)
+    b = SeqAllToAll4D.apply(cp_group, b, 2, 1)
+
+    a = a.reshape(batch_size, seq_len, self.num_v_heads, 1)
+    a = SeqAllToAll4D.apply(cp_group, a, 2, 1)
+
+    # Merge Q/K/V for conv1d (conv1d requires merged qkv)
+    q_flat = q_proj.reshape(batch_size, full_seq_len, self.key_dim // cp_size)
+    k_flat = k_proj.reshape(batch_size, full_seq_len, self.key_dim // cp_size)
+    v_flat = v_proj.reshape(batch_size, full_seq_len, self.value_dim // cp_size)
+    qkv = torch.cat([q_flat, k_flat, v_flat], dim=-1)  # [B, S, (key_dim*2+value_dim)/cp]
+
+    # Conv1d in HP layout with CP-aware weight slicing
+    mixed_qkv = qkv.transpose(1, 2).contiguous()  # [B, conv_dim/cp, S]
+    conv1d_weight = get_parameter_local_cp(
+        self.conv1d.weight,
+        dim=0,
+        cp_group=cp_group,
+        split_sections=[self.key_dim, self.key_dim, self.value_dim],
+    )
+    conv1d_bias = None
+    if self.conv1d.bias is not None:
+        conv1d_bias = get_parameter_local_cp(
+            self.conv1d.bias,
+            dim=0,
+            cp_group=cp_group,
+            split_sections=[self.key_dim, self.key_dim, self.value_dim],
+        )
+
+    if self.causal_conv1d_fn is not None:
+        mixed_qkv = self.causal_conv1d_fn(
+            x=mixed_qkv,
+            weight=conv1d_weight.squeeze(1),
+            bias=conv1d_bias,
+            activation=self.activation,
+            seq_idx=None,
+            **({"cu_seqlens": cu_seqlens} if cu_seqlens is not None else {}),
+        )
+    elif cu_seqlens is not None:
+        raise RuntimeError(
+            "cu_seqlens requires causal_conv1d_fn (FLA) but it is not available. "
+            "Please install flash-linear-attention for pack support."
+        )
+    else:
+        conv_out = F.conv1d(
+            input=mixed_qkv,
+            weight=conv1d_weight,
+            bias=conv1d_bias,
+            stride=self.conv1d.stride,
+            padding=self.conv1d.padding,
+            dilation=self.conv1d.dilation,
+            groups=self.conv_dim // cp_size,
+        )
+        mixed_qkv = self.act(conv_out[..., :full_seq_len])
+    mixed_qkv = mixed_qkv.transpose(1, 2).contiguous()  # [B, S, conv_dim/cp]
+
+    query, key, value = torch.split(
+        mixed_qkv,
+        [self.key_dim // cp_size, self.key_dim // cp_size, self.value_dim // cp_size],
+        dim=-1,
+    )
+    query = query.reshape(batch_size, full_seq_len, -1, self.head_k_dim)
+    key = key.reshape(batch_size, full_seq_len, -1, self.head_k_dim)
+    value = value.reshape(batch_size, full_seq_len, -1, self.head_v_dim)
+
+    if self.num_v_heads // self.num_k_heads > 1:
+        repeat_factor = self.num_v_heads // self.num_k_heads
+        query = query.repeat_interleave(repeat_factor, dim=2)
+        key = key.repeat_interleave(repeat_factor, dim=2)
+
+    gate = z  # [B, S, num_v_heads/cp, head_v_dim]
+    beta = b.squeeze(-1)  # [B, S, num_v_heads/cp]
+    alpha = a.squeeze(-1)  # [B, S, num_v_heads/cp]
+
+    query = query.contiguous()
+    key = key.contiguous()
+    value = value.contiguous()
+    gate = gate.contiguous()
+    beta = beta.contiguous()
+    alpha = alpha.contiguous()
+
+    A_log_local = get_parameter_local_cp(self.A_log, dim=0, cp_group=cp_group)
+    dt_bias_local = get_parameter_local_cp(self.dt_bias, dim=0, cp_group=cp_group)
+    g = -A_log_local.float().exp() * F.softplus(alpha.float() + dt_bias_local)
+    beta_final = beta.sigmoid()
+
+    # Gated delta rule in HP layout (needs full sequence)
+    core_attn_out, _ = self.chunk_gated_delta_rule(
+        query,
+        key,
+        value,
+        g=g,
+        beta=beta_final,
+        initial_state=None,
+        output_final_state=False,
+        use_qk_l2norm_in_kernel=True,
+        **({"cu_seqlens": cu_seqlens} if cu_seqlens is not None else {}),
+    )
+
+    z_shape_og = gate.shape
+    core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
+    z_flat = gate.reshape(-1, gate.shape[-1])
+    core_attn_out = self.norm(core_attn_out, z_flat)
+    core_attn_out = core_attn_out.reshape(z_shape_og)  # [B, S, num_v_heads/cp, head_v_dim]
+
+    # HP->CP all_to_all: scatter seq (dim=1), gather heads (dim=2)
+    # [B, S, num_v_heads/cp, head_v_dim] -> [B, S/cp, num_v_heads, head_v_dim]
+    norm_out = SeqAllToAll4D.apply(cp_group, core_attn_out, 1, 2)
+    norm_out = norm_out.reshape(batch_size, seq_len, -1)
+
+    # Output projection in CP layout
+    output = self.out_proj(norm_out)
+    return output

src/llamafactory/v1/plugins/model_plugins/parallelization/sequence_parallel.py

@@ -24,6 +24,7 @@ from ....accelerator.interface import Dim, DistributedInterface
 from ....utils import logging
 from ....utils.plugin import BasePlugin
 from ....utils.types import ModelOutput
+from .gdn_attention import _get_gdn_module, gdn_forward_with_cp, is_gdn_layer
 from .ulysses import (
     UlyssesAttention,
     get_ulysses_sequence_parallel_group,
@@ -127,6 +128,20 @@ def apply_sequence_parallel(model, cp_size: int):
         except (AttributeError, TypeError):
             continue
 
+    # Register GDN forward for CP support
+    if cp_size > 1:
+        replaced_modules = set()
+        for name, module in model.named_modules():
+            if is_gdn_layer(module):
+                gdn_module = _get_gdn_module(module)
+                if id(gdn_module) in replaced_modules:
+                    continue
+                replaced_modules.add(id(gdn_module))
+                gdn_module.original_forward = gdn_module.forward
+                gdn_module.forward = gdn_forward_with_cp.__get__(gdn_module, type(gdn_module))
+                gdn_name = name if gdn_module is module else f"{name}.linear_attn"
+                logger.info_rank0(f"Replaced GDN forward in {gdn_name} with gdn_forward_with_cp for context parallel.")
+
 
 def padding_and_split_data(data, device_mesh=None):
     if device_mesh is not None:

tests_v1/core/utils/test_batching.py

@@ -566,12 +566,11 @@ def test_drop_unsupervised_samples():
 
     gen = SimpleNamespace(cutoff_len=4, _warned_truncation=False)
     samples = [
-        _s([0.0, 0.0, 1.0, 1.0]),             # fits cutoff (len 4), supervised -> kept
-        _s([0.0, 0.0, 0.0, 0.0, 1.0, 1.0]),   # len 6 > 4, supervision only beyond cutoff -> dropped
-        _s([1.0, 1.0]),                        # short, fully supervised -> kept
-        _s([0.0, 0.0, 1.0, 1.0, 1.0, 1.0]),   # len 6 > 4 but supervision within cutoff -> kept
+        _s([0.0, 0.0, 1.0, 1.0]),  # fits cutoff (len 4), supervised -> kept
+        _s([0.0, 0.0, 0.0, 0.0, 1.0, 1.0]),  # len 6 > 4, supervision only beyond cutoff -> dropped
+        _s([1.0, 1.0]),  # short, fully supervised -> kept
+        _s([0.0, 0.0, 1.0, 1.0, 1.0, 1.0]),  # len 6 > 4 but supervision within cutoff -> kept
     ]
     kept = BatchGenerator._drop_unsupervised(gen, samples)
     assert kept == [samples[0], samples[2], samples[3]]
     assert gen._warned_truncation is True
-

tests_v1/plugins/data_plugins/test_converter.py

@@ -259,4 +259,3 @@ def test_pair_converter(num_samples: int):
             ],
         }
         assert data_engine[index] == {"_dataset_name": "tiny_dataset", **expected_data}
-