GLM-4-0414 and GLM-4.1V Code Refactor (#12117)

python/sglang/srt/layers/rotary_embedding.py

@@ -1070,6 +1070,7 @@ def _triton_mrope_forward(
     mrope_section_h: tl.constexpr,
     mrope_section_w: tl.constexpr,
     is_interleaved: tl.constexpr,
+    is_neox_style: tl.constexpr,
 ):
     # Adapted from
     # https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/qwen2vl_mrope.py
@@ -1124,51 +1125,99 @@ def _triton_mrope_forward(
     # program instance (i.e. for the current token) separately
     # ####################################################################
     # left half of the head
-    first_half_q_offsets = (
-        tl.arange(0, pad_n_qh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
-    )
-    first_half_k_offsets = (
-        tl.arange(0, pad_n_kh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
-    )
-    first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & (
-        tl.arange(0, pad_hd // 2)[None, :] < rd // 2
-    )
-    first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & (
-        tl.arange(0, pad_hd // 2)[None, :] < rd // 2
-    )
+    if is_neox_style:
+        first_half_q_offsets = (
+            tl.arange(0, pad_n_qh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+        )
+        first_half_k_offsets = (
+            tl.arange(0, pad_n_kh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+        )
+        first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & (
+            tl.arange(0, pad_hd // 2)[None, :] < rd // 2
+        )
+        first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & (
+            tl.arange(0, pad_hd // 2)[None, :] < rd // 2
+        )
 
-    q_tile_1 = tl.load(q_ptr + first_half_q_offsets, mask=first_q_mask, other=0).to(
-        sin_row.dtype
-    )
-    k_tile_1 = tl.load(k_ptr + first_half_k_offsets, mask=first_k_mask, other=0).to(
-        sin_row.dtype
-    )
+        q_tile_1 = tl.load(q_ptr + first_half_q_offsets, mask=first_q_mask, other=0).to(
+            sin_row.dtype
+        )
+        k_tile_1 = tl.load(k_ptr + first_half_k_offsets, mask=first_k_mask, other=0).to(
+            sin_row.dtype
+        )
 
-    # right half of the head
-    second_half_q_offsets = first_half_q_offsets + (rd // 2)
-    second_half_k_offsets = first_half_k_offsets + (rd // 2)
-    second_q_mask = first_q_mask
-    second_k_mask = first_k_mask
+        # right half of the head
+        second_half_q_offsets = first_half_q_offsets + (rd // 2)
+        second_half_k_offsets = first_half_k_offsets + (rd // 2)
+        second_q_mask = first_q_mask
+        second_k_mask = first_k_mask
+
+        q_tile_2 = tl.load(
+            q_ptr + second_half_q_offsets, mask=second_q_mask, other=0
+        ).to(sin_row.dtype)
+        k_tile_2 = tl.load(
+            k_ptr + second_half_k_offsets, mask=second_k_mask, other=0
+        ).to(sin_row.dtype)
+
+        # y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]
+        # Since cos and sin are now half-size,
+        # we use the same cos_row and sin_row for both halves
+        new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
+        tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
+        tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+        new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
+        tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
+        tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+    else:
+        base_q = tl.arange(0, pad_n_qh)[:, None] * hd
+        base_k = tl.arange(0, pad_n_kh)[:, None] * hd
+        even_idx = 2 * tl.arange(0, pad_hd // 2)[None, :]
+        odd_idx = even_idx + 1
+
+        even_q_offsets = base_q + even_idx
+        odd_q_offsets = base_q + odd_idx
+        even_k_offsets = base_k + even_idx
+        odd_k_offsets = base_k + odd_idx
+
+        idx_mask = tl.arange(0, pad_hd // 2)[None, :] < (rd // 2)
+        qn_mask = tl.arange(0, pad_n_qh)[:, None] < n_qh
+        kn_mask = tl.arange(0, pad_n_kh)[:, None] < n_kh
+
+        even_q_mask = qn_mask & idx_mask
+        odd_q_mask = qn_mask & idx_mask
+        even_k_mask = kn_mask & idx_mask
+        odd_k_mask = kn_mask & idx_mask
+
+        q_tile_1 = tl.load(q_ptr + even_q_offsets, mask=even_q_mask, other=0).to(
+            sin_row.dtype
+        )
+        k_tile_1 = tl.load(k_ptr + even_k_offsets, mask=even_k_mask, other=0).to(
+            sin_row.dtype
+        )
 
-    q_tile_2 = tl.load(q_ptr + second_half_q_offsets, mask=second_q_mask, other=0).to(
-        sin_row.dtype
-    )
-    k_tile_2 = tl.load(k_ptr + second_half_k_offsets, mask=second_k_mask, other=0).to(
-        sin_row.dtype
-    )
+        q_tile_2 = tl.load(q_ptr + odd_q_offsets, mask=odd_q_mask, other=0).to(
+            sin_row.dtype
+        )
+        k_tile_2 = tl.load(k_ptr + odd_k_offsets, mask=odd_k_mask, other=0).to(
+            sin_row.dtype
+        )
 
-    # y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]
-    # Since cos and sin are now half-size,
-    # we use the same cos_row and sin_row for both halves
-    new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
-    tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
-    new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
-    tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+        # y = [x_even, x_odd] * [cos, cos] + [-x_odd, x_even] * [sin, sin]
+        # NeoX-style rotary embedding:
+        # Each (even, odd) channel pair forms one rotation arm.
+        # cos_row and sin_row each have length rd//2, shared across all (even, odd) pairs.
+        new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
+        tl.store(q_ptr + even_q_offsets, new_q_tile_1, mask=even_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
+        tl.store(q_ptr + odd_q_offsets, new_q_tile_2, mask=odd_q_mask)
 
-    new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
-    tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
-    new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
-    tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+        new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
+        tl.store(k_ptr + even_k_offsets, new_k_tile_1, mask=even_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
+        tl.store(k_ptr + odd_k_offsets, new_k_tile_2, mask=odd_k_mask)
 
 
 def triton_mrope(
@@ -1180,6 +1229,7 @@ def triton_mrope(
     head_size: int,
     rotary_dim: int,
     mrope_interleaved: bool,
+    is_neox_style: bool,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """The mrope triton kernel.
 
@@ -1230,6 +1280,7 @@ def triton_mrope(
         mrope_section[1],
         mrope_section[2],
         mrope_interleaved,
+        is_neox_style,
     )
     return q, k
 
@@ -1400,6 +1451,7 @@ class MRotaryEmbedding(RotaryEmbedding):
                 self.head_size,
                 self.rotary_dim,
                 self.mrope_interleaved,
+                self.is_neox_style,
             )
 
             return q.reshape(query_shape), k.reshape(key_shape)

python/sglang/srt/models/glm4v.py

+# Copyright 2023-2024 SGLang 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.
+# ==============================================================================
+
+# Modeling from:
+# ./llama.py and
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4v/modular_glm4v.py
+"""Inference-only GLM-4.1V model compatible with HuggingFace weights."""
+
 import logging
-from functools import lru_cache, partial
+from functools import lru_cache
 from typing import Iterable, List, Optional, Tuple
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from einops import rearrange
 from transformers.models.glm4v.configuration_glm4v import Glm4vConfig, Glm4vVisionConfig
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.attention import vision_utils
-from sglang.srt.layers.dp_attention import get_attention_tp_size
+from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     ColumnParallelLinear,
@@ -20,13 +40,14 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.pooler import Pooler, PoolingType
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
-from sglang.srt.managers.schedule_batch import MultimodalDataItem
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternMultimodalTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.glm4 import Glm4Model
-from sglang.srt.models.qwen2_5_vl import (
-    Qwen2_5_VisionBlock,
-    Qwen2_5_VLForConditionalGeneration,
-)
 from sglang.srt.utils import add_prefix
 from sglang.srt.utils.hf_transformers_utils import get_processor
 
@@ -56,7 +77,7 @@ class Glm4vVisionMLP(nn.Module):
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
             input_size=in_features,
-            output_sizes=[hidden_features] * 2,
+            output_sizes=[hidden_features] * 2,  # [gate_proj, up_proj]
             bias=bias,
             quant_config=quant_config,
             prefix=add_prefix("gate_up_proj", prefix),
@@ -77,34 +98,95 @@ class Glm4vVisionMLP(nn.Module):
         return x
 
 
-class Glm4vVisionBlock(Qwen2_5_VisionBlock):
+class Glm4vVisionBlock(nn.Module):
     def __init__(
         self,
-        config: Glm4vVisionConfig,
-        norm_layer: Optional[nn.Module] = None,
+        dim: int,
+        intermediate_dim: int,
+        num_heads: int,
+        attn_implementation: Optional[str] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        num_dummy_heads: int = 0,
+        rms_norm_eps: float = 1e-5,
     ) -> None:
-        super().__init__(
-            dim=config.hidden_size,
-            intermediate_dim=config.out_hidden_size,
-            num_heads=config.num_heads,
-            hidden_act=config.hidden_act,
-            norm_layer=norm_layer,
+        super().__init__()
+        self.norm1 = RMSNorm(dim, eps=rms_norm_eps)
+        self.norm2 = RMSNorm(dim, eps=rms_norm_eps)
+
+        if attn_implementation is None:
+            softmax_in_single_precision = False
+            qkv_backend = None
+            flatten_batch = True
+        elif attn_implementation == "sdpa":
+            softmax_in_single_precision = False
+            qkv_backend = "sdpa"
+            flatten_batch = True
+        elif attn_implementation == "flash_attention_2":
+            softmax_in_single_precision = False
+            qkv_backend = "triton_attn"
+            flatten_batch = True
+        elif attn_implementation == "eager":
+            softmax_in_single_precision = True
+            qkv_backend = "sdpa"
+            flatten_batch = True
+        elif attn_implementation == "flash_attention_3":
+            softmax_in_single_precision = False
+            qkv_backend = "fa3"
+            flatten_batch = True
+
+        self.attn = VisionAttention(
+            embed_dim=dim,
+            num_heads=num_heads,
+            projection_size=dim,
+            use_qkv_parallel=True,
+            rotary_embed="normal",
+            proj_bias=True,
+            qkv_backend=qkv_backend,
+            softmax_in_single_precision=softmax_in_single_precision,
+            flatten_batch=flatten_batch,
             quant_config=quant_config,
-            prefix=prefix,
-            num_dummy_heads=config.num_dummy_heads,
-            rms_norm_eps=config.rms_norm_eps,
+            prefix=add_prefix("attn", prefix),
+            num_dummy_heads=num_dummy_heads,
         )
-
         self.mlp = Glm4vVisionMLP(
-            config.hidden_size,
-            config.out_hidden_size,
-            bias=False,
+            dim,
+            intermediate_dim,
             quant_config=quant_config,
             prefix=add_prefix("mlp", prefix),
         )
 
+    def forward(
+        self,
+        x: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: torch.Tensor,
+    ) -> torch.Tensor:
+        S, B, H = x.shape
+        # norm1: flatten to 2D -> [S*B, H], then reshape back
+        x2d = x.reshape(-1, H)
+        hidden_states = self.norm1(x2d).reshape(S, B, H)
+
+        # Attention expects [B, S, H]
+        hidden_states = rearrange(hidden_states, "s b h -> b s h")
+        attn = self.attn(
+            hidden_states,
+            cu_seqlens=cu_seqlens,
+            position_embeddings=position_embeddings,
+        )
+        attn = rearrange(attn, "b s h -> s b h")
+
+        # norm2 with fused residual-add: also 2D
+        attn2d = attn.reshape(-1, H)
+        x_norm_2d, x_after_add_2d = self.norm2(x2d, residual=attn2d)
+        x_norm = x_norm_2d.reshape(S, B, H)
+        x_after_add = x_after_add_2d.reshape(S, B, H)
+
+        # MLP and final residual
+        mlp_out = self.mlp(x_norm)
+        x = x_after_add + mlp_out
+        return x
+
 
 class Glm4vVisionPatchEmbed(nn.Module):
     def __init__(
@@ -320,7 +402,6 @@ class Glm4vVisionModel(nn.Module):
     def __init__(
         self,
         vision_config: Glm4vVisionConfig,
-        norm_eps: float = 1e-6,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
@@ -344,17 +425,18 @@ class Glm4vVisionModel(nn.Module):
             hidden_size=self.hidden_size,
         )
 
-        norm_layer = partial(Glm4vRMSNorm, eps=norm_eps)
         head_dim = self.hidden_size // self.num_heads
         self.rotary_pos_emb = Glm4vVisionRotaryEmbedding(head_dim // 2)
 
         self.blocks = nn.ModuleList(
             [
                 Glm4vVisionBlock(
-                    config=vision_config,
-                    norm_layer=norm_layer,
+                    dim=self.hidden_size,
+                    intermediate_dim=self.out_hidden_size,
+                    num_heads=self.num_heads,
                     quant_config=quant_config,
                     prefix=add_prefix(f"blocks.{layer_idx}", prefix),
+                    rms_norm_eps=vision_config.rms_norm_eps,
                 )
                 for layer_idx in range(depth)
             ]
@@ -461,29 +543,30 @@ class Glm4vVisionModel(nn.Module):
         return x
 
 
-class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
+class Glm4vForConditionalGeneration(nn.Module):
     def __init__(
         self,
         config: Glm4vConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
-        nn.Module.__init__(self)
+        super().__init__()
 
         self.config = config
-        vision_utils.update_vit_attn_dummy_heads_config(self.config)
-        self.model = Glm4Model(
-            config,
-            quant_config,
-            prefix=add_prefix("model", prefix),
-        )
         self.visual = Glm4vVisionModel(
             config.vision_config,
-            norm_eps=getattr(config, "rms_norm_eps", 1e-5),
             quant_config=quant_config,
             prefix=add_prefix("visual", prefix),
         )
 
+        vision_utils.update_vit_attn_dummy_heads_config(self.config)
+
+        self.model = Glm4Model(
+            config,
+            quant_config=quant_config,
+            prefix=add_prefix("model", prefix),
+        )
+
         if config.tie_word_embeddings:
             self.lm_head = self.model.embed_tokens
         else:
@@ -494,13 +577,18 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
                 prefix=add_prefix("lm_head", prefix),
             )
 
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
+
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
-        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
 
         # For EAGLE3 support
         self.capture_aux_hidden_states = False
 
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
         pixel_values = torch.cat(
             [item.feature.squeeze(0) for item in items], dim=0
@@ -542,20 +630,60 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         video_embeds = torch.split(video_embeds, split_sizes)
         return torch.cat(video_embeds)
 
-    def _update_hf_config(self):
-        """update hf config to ensure vision attention num_attention_heads is divisible by tp_size"""
-        tp_size = get_attention_tp_size()
-        num_heads = self.config.vision_config.num_heads
-        head_dim = self.config.vision_config.hidden_size // num_heads
-        num_dummy_heads = 0
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
 
-        if num_heads % tp_size != 0:
-            num_dummy_heads = (
-                (num_heads + tp_size - 1) // tp_size
-            ) * tp_size - num_heads
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        get_embedding: bool = False,
+    ):
+        """Run forward pass for GLM-4.1V.
+
+        Args:
+            input_ids: Flattened (concatenated) input_ids corresponding to a
+                batch.
+            positions: Flattened (concatenated) position ids corresponding to a
+                batch.
+                **NOTE**: If mrope is enabled (default setting for GLM-4.1V
+                opensource models), the shape will be `(3, seq_len)`,
+                otherwise it will be `(seq_len,).
+                (Use input_metadata.mrope_positions to replace it)
+        """
+        if self.is_mrope_enabled:
+            positions = forward_batch.mrope_positions
+
+        if not (
+            forward_batch.forward_mode.is_decode()
+            or not forward_batch.contains_image_inputs()
+        ):
+            if self.is_mrope_enabled:
+                assert positions.ndim == 2 and positions.size(0) == 3, (
+                    "multimodal section rotary embedding requires "
+                    f"(3, seq_len) positions, but got {positions.size()}"
+                )
+
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.model,
+            multimodal_model=self,
+            positions=positions,
+        )
 
-        setattr(self.config.vision_config, "head_dim", head_dim)
-        setattr(self.config.vision_config, "num_dummy_heads", num_dummy_heads)
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
+        if not get_embedding:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch, aux_hidden_states
+            )
+        else:
+            return self.pooler(hidden_states, forward_batch)
 
     def _pad_vit_attn_dummy_heads(self, name: str, loaded_weight: torch.Tensor):
         """pad attn qkv weights for dummy heads"""
@@ -598,13 +726,12 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
         ]
         params_dict = dict(self.named_parameters(remove_duplicate=False))
         for name, loaded_weight in weights:
-            if "language_model." in name:
-                name = name.replace("language_model.", "")
-            if "model.visual." in name:
-                name = name.replace("model.visual.", "visual.")
-
             if "rotary_emb.inv_freq" in name:
                 continue
+            if "language_model" in name:
+                name = name.replace(r"model.language_model.", r"model.")
+            if "model.visual." in name:
+                name = name.replace("model.visual.", "visual.")
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
@@ -639,5 +766,19 @@ class Glm4vForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
                     )
                 weight_loader(param, loaded_weight)
 
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        self.model.embed_tokens.weight = embed
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            del self.lm_head.weight
+            self.lm_head.weight = head
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
 
 EntryClass = [Glm4vForConditionalGeneration]

python/sglang/srt/models/glm4v_moe.py

@@ -53,7 +53,6 @@ class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
         )
         self.visual = Glm4vVisionModel(
             config.vision_config,
-            norm_eps=getattr(config, "rms_norm_eps", 1e-5),
             quant_config=quant_config,
             prefix=add_prefix("visual", prefix),
         )