# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/glm5_next/modular_glm5_next.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_glm5_next.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # Copyright 2026 the HuggingFace Team. All rights reserved. # # 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 math from collections.abc import Callable from itertools import accumulate import torch import torch.nn as nn import torch.nn.functional as F from torch.nn import LayerNorm from ... import initialization as init from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache from ...generation import GenerationMixin from ...integrations import ( use_experts_implementation, use_kernel_forward_from_hub, use_kernel_func_from_hub_with_fallback, use_kernelized_func, ) from ...integrations.accelerate import force_accelerate_hooks from ...masking_utils import create_recurrent_attention_mask from ...modeling_flash_attention_utils import FlashAttentionKwargs from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import ( BaseModelOutputWithPast, BaseModelOutputWithPooling, MoeCausalLMOutputWithPast, MoeModelOutputWithPast, ) from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check from ...utils.generic import ( accepts_precomputed_kwargs, get_max_seqlen, is_flash_attention_requested, merge_with_config_defaults, ) from ...utils.output_capturing import OutputRecorder, capture_outputs from ...vision_utils import get_vision_attention_seqlens, get_vision_position_ids from .configuration_glm5_next import Glm5NextConfig, Glm5NextTextConfig, Glm5NextVisionConfig @use_kernel_forward_from_hub("RMSNorm") class Glm5NextTextRMSNorm(nn.Module): def __init__(self, hidden_size, eps: float = 1e-6) -> None: """ Glm5NextTextRMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" class Glm5NextTextMLP(nn.Module): def __init__(self, config, intermediate_size=None): super().__init__() self.config = config self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) self.act_fn = ACT2FN[config.hidden_act] self.swiglu_limit = config.swiglu_limit def forward(self, x): gate = self.gate_proj(x) up = self.up_proj(x) # Key difference using clamping gate = gate.clamp(min=None, max=self.swiglu_limit) up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit) return self.down_proj(self.act_fn(gate) * up) @use_experts_implementation class Glm5NextTextExperts(nn.Module): """Collection of expert weights stored as 3D tensors.""" def __init__(self, config): super().__init__() self.num_experts = config.num_local_experts self.hidden_dim = config.hidden_size self.intermediate_dim = config.moe_intermediate_size self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim)) self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim)) self.swiglu_limit = config.swiglu_limit def forward( self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor ) -> torch.Tensor: final = torch.zeros_like(hidden_states) with torch.no_grad(): mask = F.one_hot(top_k_index, num_classes=self.num_experts + 1).permute(2, 1, 0) hit = torch.greater(mask.sum(dim=(-1, -2)), 0).nonzero() for expert_idx in hit: expert_idx = expert_idx[0] if expert_idx == self.num_experts: continue top_k_pos, token_idx = torch.where(mask[expert_idx]) current = self._apply_gate(F.linear(hidden_states[token_idx], self.gate_up_proj[expert_idx])) current = F.linear(current, self.down_proj[expert_idx]) * top_k_weights[token_idx, top_k_pos, None] final.index_add_(0, token_idx, current.to(final.dtype)) return final def _apply_gate(self, gate_up: torch.Tensor) -> torch.Tensor: gate, up = gate_up.chunk(2, dim=-1) gate = gate.clamp(min=None, max=self.swiglu_limit) up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit) # Simple swiglu instead of alpha return F.silu(gate) * up class Glm5NextTextTopkRouter(nn.Module): def __init__(self, config: Glm5NextTextConfig): super().__init__() self.top_k = config.num_experts_per_tok self.num_experts = config.num_local_experts self.hidden_dim = config.hidden_size self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim)) self.routed_scaling_factor = config.routed_scaling_factor self.num_group = config.n_group self.topk_group = config.topk_group self.norm_topk_prob = config.norm_topk_prob self.e_score_correction_bias = nn.Buffer(torch.zeros(self.num_experts)) def forward(self, hidden_states): hidden_states = hidden_states.view(-1, self.hidden_dim) router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32)) scores = router_logits.sigmoid() scores_for_choice = scores + self.e_score_correction_bias group_scores = ( scores_for_choice.view(-1, self.num_group, self.num_experts // self.num_group) .topk(2, dim=-1)[0] .sum(dim=-1) ) group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1] group_mask = torch.zeros_like(group_scores) group_mask.scatter_(1, group_idx, 1) score_mask = ( group_mask.unsqueeze(-1) .expand(-1, self.num_group, self.num_experts // self.num_group) .reshape(-1, self.num_experts) ) scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), float("-inf")) topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1] topk_weights = scores.gather(1, topk_indices) if self.norm_topk_prob: denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20 topk_weights /= denominator topk_weights = topk_weights * self.routed_scaling_factor return router_logits, topk_weights, topk_indices class Glm5NextTextMoE(nn.Module): """ A mixed expert module containing shared experts. """ def __init__(self, config: Glm5NextConfig): super().__init__() self.config = config self.experts = Glm5NextTextExperts(config) self.gate = Glm5NextTextTopkRouter(config) self.shared_experts = Glm5NextTextMLP( config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: residuals = hidden_states orig_shape = hidden_states.shape _, topk_weights, topk_indices = self.gate(hidden_states) hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape) hidden_states = hidden_states + self.shared_experts(residuals) return hidden_states class Glm5NextTextUnweightedRMSNorm(nn.Module): def __init__(self, eps: float = 1.0e-6): super().__init__() self.eps = eps def forward(self, x: torch.Tensor) -> torch.Tensor: return x * torch.rsqrt(x.float().square().mean(-1, keepdim=True) + self.eps).to(x.dtype) class Glm5NextTextHyperConnection(nn.Module): r""" A module to implement manifold-constrained Hyper-Connections (mHC) (Xie et al., 2026) which strengthens the conventional residual connections between adjacent Transformer blocks. When using mHC, each token is projected onto `hc_mult` streams, so the shape of decoder layer inputs changes from [batch_size, sequence_length, hidden_size] to [batch_size, sequence_length, hc_mult, hidden_size]. To keep the same input shape for attention or MLP blocks, the streams are collapsed into one upon entering a block, and expanded back into `hc_mult` streams upon exiting. There is also a weighted residual connection between the input and output streams. The weights used for collapsing (pre), expanding (post) and mixing (comb) are computed from the `hc_mult` input streams through a learned projection (plus Sinkhorn-Knopp algorithm for the comb weight). The diagram below shows the flow of the mHC streams (B = batch_size, S = seq_length, N = hc_mult, D = hidden_size): ┌───────────────────┐ N input streams ────│ FLATTEN + PROJECT |────> (pre, post, comb) weights [B, S, N, D] └───────────────────┘ ([B, S, N], [B, S, N], [B, S, N, N]) │ │ │ ╭──────────────────┴─┼─┼──────────────────╮ │ ╭──────────────────┴─┼────────────────╮ │ │ │ ╭──────────────────┴──────────────╮ │ │ │ │ │ │ │ │ ┌─────────────────┐ │ │ │ │ COLLAPSE (pre) │ │ │ │ └─────────────────┘ │ │ │ │ │ │ │ Block input │ │ │ [B, S, D] │ │ │ │ │ │ │ ┌─────────────────┐ ┌─────────────┐ │ ATTN or MLP │ │ MIX (comb) │ └─────────────────┘ └─────────────┘ │ │ │ │ Block output │ │ │ [B, S, D] │ │ │ │ │ │ │ ┌─────────────────┐ │ │ │ │ EXPAND (post) │ │ │ │ └─────────────────┘ │ │ │ │ │ │ │ │ │ N expanded streams N residual streams [B, S, N, D] [B, S, N, D] │ │ │ │ │ │ │ │ ╰───────────┌─────────┐───────────╯ │ │ │ ╰─────────────│ ADD │─────────────╯ │ ╰───────────────└─────────┘───────────────╯ │ │ │ ▼ ▼ ▼ N output streams [B, S, N, D] """ def __init__(self, config: Glm5NextTextConfig): super().__init__() self.hc_mult = config.hc_mult # number of streams, referred as N below self.hc_sinkhorn_iters = config.hc_sinkhorn_iters self.hc_eps = config.hc_eps self.input_norm = Glm5NextTextUnweightedRMSNorm(eps=config.rms_norm_eps) # The mHC projects the N inputs streams into 3 weights: pre (size: N), post (size: N) and comb (size: N*N) # Hence the output size of the projection is 2 * N + N * N = (2 + N) * N. concatenated_weights_size = (2 + self.hc_mult) * self.hc_mult self.fn = nn.Parameter(torch.empty(concatenated_weights_size, self.hc_mult * config.hidden_size)) self.base = nn.Parameter(torch.empty(concatenated_weights_size)) # The mHC produces 3 outputs, each with their own scale parameter (the "pre", "post" and "comb" weights) self.scale = nn.Parameter(torch.empty(3)) def forward(self, hidden_streams: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """ Computes the weights used to mix in the `hc_mult` streams with the input and output of the next layer, which can be an attention or a MLP layer. This is done through three weights: - pre: used to collapse the `hc_mult` input streams into one, creating an input tensor for the next layer - post: used to expand the output of the next layer back into `hc_mult` streams - comb: used to mix the `hc_mult` input streams with the `hc_mult` output streams All weights are returned except "pre", which is consumed here. """ batch_size, seq_len = hidden_streams.shape[:2] hc = self.hc_mult # Flatten and norm the hidden streams flattened = hidden_streams.view(batch_size, seq_len, -1).float() flattened = self.input_norm(flattened) # Mix the streams together to infer the weight coefficients flattened = F.linear(flattened, self.fn.float()) # Split the weight coefficients pre_w, post_w, comb_w = flattened.split([hc, hc, hc * hc], dim=-1) pre_b, post_b, comb_b = self.base.split([hc, hc, hc * hc]) pre_scale, post_scale, comb_scale = self.scale.unbind(0) comb_w = comb_w.view(*comb_w.shape[:-1], hc, hc) # these are matrix weights, unlike pre or post comb_b = comb_b.view(hc, hc) # All weights are computed with a one layer perceptron. For pre and post, this is it. pre = torch.sigmoid(pre_w * pre_scale + pre_b) + self.hc_eps post = 2 * torch.sigmoid(post_w * post_scale + post_b) comb = torch.softmax(comb_w * comb_scale + comb_b, dim=-1) + self.hc_eps # The comb weight is a bit different: it dictates how the input streams (In) are added to the output streams # (Out) in this way: Mixed = In @ Comb + Out. To make sure the norm of "Mixed" does not blow up, we constrain # the comb weight to be doubly-stochastic (ie. its rows and columns must sum to 1) with a few iterations of the # Sinkhorn-Knopp algorithm, which iteratively normalizes the rows and columns to sum to 1. comb = comb / (comb.sum(dim=-2, keepdim=True) + self.hc_eps) for _ in range(self.hc_sinkhorn_iters - 1): comb = comb / (comb.sum(dim=-1, keepdim=True) + self.hc_eps) comb = comb / (comb.sum(dim=-2, keepdim=True) + self.hc_eps) # Since "pre" is meant to be used with the input streams (available here as `hidden_streams`), we collapse the # streams here and return `collapsed` tensor, which will be the input for the next attention or MLP block. collapsed = (pre.unsqueeze(-1) * hidden_streams).sum(dim=2).to(hidden_streams.dtype) return post, comb, collapsed class Glm5NextTextHyperHead(nn.Module): """Final GLM-5.3-Flash HC-stream collapse. Unlike DeepSeek-V4, this is an unweighted mean.""" def forward(self, hidden_streams: torch.Tensor) -> torch.Tensor: return hidden_streams.mean(dim=2) class Glm5NextTextForgetGate(nn.Module): def __init__(self, config: Glm5NextTextConfig): super().__init__() self.head_dim = config.linear_head_dim self.num_heads = config.linear_num_heads self.qkv_dim = self.head_dim * self.num_heads self.f_a_proj = nn.Linear(config.hidden_size, self.head_dim, bias=False) self.f_b_proj = nn.Linear(self.head_dim, self.qkv_dim, bias=False) self.dt_bias = nn.Parameter(torch.empty(self.qkv_dim)) self.A_log = nn.Parameter(torch.empty(self.num_heads)) self.safe_gate_lower_bound = config.linear_lower_bound def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: hidden_shape = (*hidden_states.shape[:2], -1, self.head_dim) forget_gate = self.f_b_proj(self.f_a_proj(hidden_states)) g = (forget_gate.float() + self.dt_bias.float().view(1, 1, -1)).view(hidden_shape) A_log = self.A_log.float().view(1, 1, self.num_heads, 1) decay_rate = torch.exp(A_log) # Safe lower bound decay if self.safe_gate_lower_bound is not None: return self.safe_gate_lower_bound * torch.sigmoid(decay_rate * g) # Softplus "log(1 + exp(x))" with uper bound restraint to avoid overflows # NOTE: Softplus for larger values (e.g. 20+), Softplus(x) == x g_softplus = torch.where(g > 20.0, g, torch.log(1.0 + torch.exp(g))) return -decay_rate * g_softplus # NOTE: the FLA package does not re-cast to `input_dtype` in its implementation, maybe we should do the same @use_kernel_forward_from_hub("RMSNormGated") class Glm5NextTextRMSNormGated(nn.Module): def __init__(self, hidden_size, eps=1e-6, **kwargs) -> None: super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps self.activation = "sigmoid" def forward(self, hidden_states, gate=None) -> torch.Tensor: input_dtype = hidden_states.dtype # Strict FP32 norm (do not downcast on the weights) hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) hidden_states = self.weight.to(torch.float32) * hidden_states # Apply gating hidden_states = hidden_states * ACT2FN[self.activation](gate.to(torch.float32)) return hidden_states.to(input_dtype) def apply_mask_to_padding_states(hidden_states, attention_mask): """ Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66 """ # NOTE: attention mask is a 2D boolean tensor if attention_mask is not None: dtype = hidden_states.dtype hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype) return hidden_states @use_kernel_func_from_hub_with_fallback("causal_conv1d_update", "causal_conv1d") def causal_conv1d_update( hidden_states: torch.Tensor, conv_state: torch.Tensor, weight: nn.Parameter, bias: nn.Parameter | None = None, activation: str | None = None, ): _, hidden_size, seq_len = hidden_states.shape state_len = conv_state.shape[-1] hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype) conv_state.copy_(hidden_states_new[:, :, -state_len:]) out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size) out = out[:, :, -seq_len:] if activation is not None: out = ACT2FN[activation](out) return out.to(hidden_states.dtype) @use_kernel_func_from_hub_with_fallback("causal_conv1d_fn", "causal_conv1d") def causal_conv1d_fn( hidden_states: torch.Tensor, weight: nn.Parameter, bias: nn.Parameter | None = None, activation: str | None = None, **kwargs, ): _, hidden_size, seq_len = hidden_states.shape padding = weight.shape[-1] - 1 out = F.conv1d( hidden_states.to(weight.dtype), weight=weight.unsqueeze(1), bias=bias, padding=padding, groups=hidden_size, )[:, :, :seq_len] if activation is not None: out = ACT2FN[activation](out) return out.to(hidden_states.dtype) def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6): """ This function is intended to align with the l2norm implementation in the FLA library. # NOTE: FLA compares against `F.normalize` but does + eps instead of max(..., eps) leading to a slight differences """ # main difference to qwen's gdn variation: intentionally use sqrt and / to match original triton inv_norm = torch.sqrt((x * x).sum(dim=dim, keepdim=True) + eps) return x / inv_norm @use_kernel_func_from_hub_with_fallback("fused_recurrent_kda", "fla") def recurrent_kimi_delta_attention( query, key, value, g, beta, initial_state, output_final_state, use_qk_l2norm_in_kernel=False, **kwargs, ): # calculations happen in float as states are more susceptible to rounding errors initial_dtype = query.dtype query, key, value, g, beta = [x.to(torch.float32) for x in (query, key, value, g, beta)] # important: FLA calculates these in fp32 so we do this after the float casts if use_qk_l2norm_in_kernel: query = l2norm(query, dim=-1, eps=1e-6) key = l2norm(key, dim=-1, eps=1e-6) # shapes and other metadata batch_size, sequence_length, num_heads, k_head_dim = key.shape v_head_dim = value.shape[-1] scale = 1 / (query.shape[-1] ** 0.5) query = query * scale core_attn_out = torch.zeros( batch_size, sequence_length, num_heads, v_head_dim, dtype=value.dtype, device=value.device ) last_recurrent_state = ( torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim, dtype=value.dtype, device=value.device) if initial_state is None else initial_state.to(value) ) # recurrent iteration for i in range(sequence_length): q_i = query[:, i] k_i = key[:, i] v_i = value[:, i] g_i = g[:, i][..., None].exp() b_i = beta[:, i][..., None] last_recurrent_state = last_recurrent_state * g_i kv_mem = (last_recurrent_state * k_i[..., None]).sum(dim=-2) delta = (v_i - kv_mem) * b_i last_recurrent_state = last_recurrent_state + k_i.unsqueeze(-1) * delta.unsqueeze(-2) core_attn_out[:, i] = (last_recurrent_state * q_i.unsqueeze(-1)).sum(dim=-2) return core_attn_out.to(initial_dtype), last_recurrent_state if output_final_state else None @use_kernel_func_from_hub_with_fallback("chunk_kda", "fla") def chunk_kimi_delta_attention( query, key, value, g, beta, chunk_size=64, initial_state=None, output_final_state=False, use_qk_l2norm_in_kernel=False, **kwargs, ): # calculations happen in float as states are more susceptible to rounding errors initial_dtype = query.dtype query, key, value, beta, g = [ x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g) ] # important: FLA calculates these in fp32 so we do this after the float casts if use_qk_l2norm_in_kernel: query = l2norm(query, dim=-1, eps=1e-6) key = l2norm(key, dim=-1, eps=1e-6) # shapes and other metadata batch_size, num_heads, sequence_length, k_head_dim = key.shape v_head_dim = value.shape[-1] scale = 1 / (query.shape[-1] ** 0.5) pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size total_sequence_length = sequence_length + pad_size # prepare all the relevant input query = F.pad(query, (0, 0, 0, pad_size)) * scale key = F.pad(key, (0, 0, 0, pad_size)) value = F.pad(value, (0, 0, 0, pad_size)) g = F.pad(g, (0, 0, 0, pad_size)) beta = F.pad(beta, (0, pad_size)) v_beta = value * beta.unsqueeze(-1) k_beta = key * beta.unsqueeze(-1) # reshape to chunks query, key, value, g, k_beta, v_beta = [ x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1]) for x in (query, key, value, g, k_beta, v_beta) ] beta = beta.reshape(beta.shape[0], beta.shape[1], -1, chunk_size) # Intra chunk # Main difference to GDN is the per head application of `g` which was broadcasted across heads instead g = g.cumsum(dim=-2) mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0) strict_mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1) decay_mask = (g.unsqueeze(-2) - g.unsqueeze(-3)).masked_fill(strict_mask[..., None], float("-inf")).exp().float() attn = -(k_beta.unsqueeze(-2) * key.unsqueeze(-3) * decay_mask).sum(dim=-1).masked_fill(mask, 0) for i in range(1, chunk_size): row = attn[..., i, :i].clone() sub = attn[..., :i, :i].clone() attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2) attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device) value = attn @ v_beta k_cumdecay = attn @ (k_beta * g.exp()) last_recurrent_state = ( torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim, dtype=value.dtype, device=value.device) if initial_state is None else initial_state.to(value) ) core_attn_out = torch.zeros_like(value) mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1) for i in range(total_sequence_length // chunk_size): q_i = query[:, :, i] k_i = key[:, :, i] v_i = value[:, :, i] g_i = g[:, :, i] # Inter chunk attn_inter = (q_i * g_i.exp()) @ last_recurrent_state # Intra chunk attn_intra = (q_i.unsqueeze(-2) * k_i.unsqueeze(-3) * decay_mask[:, :, i]).sum(dim=-1).masked_fill(mask, 0) # New update rule v_prime = k_cumdecay[:, :, i] @ last_recurrent_state v_new = v_i - v_prime core_attn_out[:, :, i] = attn_inter + attn_intra @ v_new last_recurrent_state = ( last_recurrent_state * g_i[:, :, -1].exp().unsqueeze(-1) + (k_i * (g_i[:, :, -1:] - g_i).exp()).transpose(-1, -2) @ v_new ) if not output_final_state: last_recurrent_state = None core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1, core_attn_out.shape[-1]) core_attn_out = core_attn_out[:, :, :sequence_length] core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype) return core_attn_out, last_recurrent_state @use_kernelized_func( [chunk_kimi_delta_attention, recurrent_kimi_delta_attention, causal_conv1d_fn, causal_conv1d_update] ) class Glm5NextTextLinearAttention(nn.Module): """Kimi-style KDA (Kimi Linear Attention) for GLM-5.3-Flash.""" def __init__( self, config: Glm5NextTextConfig, layer_idx: int, ): super().__init__() self.hidden_size = config.hidden_size self.num_heads = config.linear_num_heads self.head_dim = config.linear_head_dim self.qkv_dim = self.head_dim * self.num_heads self.conv_kernel_size = config.linear_conv_kernel_dim self.layer_idx = layer_idx self.activation = config.hidden_act self.layer_norm_epsilon = config.rms_norm_eps self.q_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.qkv_dim, bias=False) self.conv_dim = self.qkv_dim * 3 self.conv1d = nn.Conv1d( in_channels=self.conv_dim, out_channels=self.conv_dim, bias=False, kernel_size=self.conv_kernel_size, groups=self.conv_dim, padding=self.conv_kernel_size - 1, ) self.forget_gate = Glm5NextTextForgetGate(config) self.b_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False) self.g_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False) self.g_b_proj = nn.Linear(self.head_dim, self.qkv_dim, bias=False) self.o_norm = Glm5NextTextRMSNormGated(self.head_dim, eps=self.layer_norm_epsilon) self.o_proj = nn.Linear(self.qkv_dim, self.hidden_size, bias=False) self.layer_type = config.layer_types[layer_idx] @force_accelerate_hooks("conv1d") def forward( self, hidden_states: torch.Tensor, cache_params: Cache | None = None, attention_mask: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ): # Zero out padding hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask) # Set up dimensions for reshapes later batch_size, seq_len = hidden_states.shape[:2] hidden_shape = (batch_size, seq_len, -1, self.head_dim) mixed_qkv = torch.cat( [ self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(hidden_states), ], dim=-1, ).transpose(1, 2) # Acts for normal prefill but also for multi-token prefill continue use_precomputed_states = cache_params is not None and cache_params.has_previous_state(self.layer_idx) if use_precomputed_states: conv_state = cache_params.layers[self.layer_idx].conv_states[0] recurrent_state = cache_params.layers[self.layer_idx].recurrent_states[0] # Single token decode path if use_precomputed_states and seq_len == 1: mixed_qkv = causal_conv1d_update( mixed_qkv, conv_state, weight=self.conv1d.weight.squeeze(1), bias=self.conv1d.bias, activation=self.activation, ) # Multi token prefill or simple "full" prefill else: # Concatenated state for prefill if cache_params is not None: mixed_qkv = cache_params.update_conv_state( mixed_qkv, self.layer_idx, conv_kernel_size=self.conv_kernel_size ) mixed_qkv = causal_conv1d_fn( mixed_qkv, weight=self.conv1d.weight.squeeze(1), bias=self.conv1d.bias, activation=self.activation, **kwargs, ) # Cut out any tail mixed_qkv = mixed_qkv[:, :, -seq_len:] query, key, value = torch.split( mixed_qkv.transpose(1, 2), [self.qkv_dim] * 3, dim=-1, ) query = query.view(hidden_shape) key = key.view(hidden_shape) value = value.view(hidden_shape) # Forget gate and input gate g = self.forget_gate(hidden_states) beta = torch.sigmoid(self.b_proj(hidden_states)) # KDA if use_precomputed_states and seq_len == 1: core_attn_out, last_recurrent_state = recurrent_kimi_delta_attention( query, key, value, g=g, beta=beta, initial_state=recurrent_state, output_final_state=cache_params is not None, use_qk_l2norm_in_kernel=True, **kwargs, ) else: core_attn_out, last_recurrent_state = chunk_kimi_delta_attention( query, key, value, g=g, beta=beta, initial_state=recurrent_state if use_precomputed_states else None, output_final_state=cache_params is not None, use_qk_l2norm_in_kernel=True, **kwargs, ) if cache_params is not None: cache_params.update_recurrent_state(last_recurrent_state.to(torch.float32), self.layer_idx) # Final gated norm and proj gate = self.g_b_proj(self.g_a_proj(hidden_states)).view(hidden_shape) output = self.o_norm(core_attn_out, gate).reshape(batch_size, seq_len, -1) output = self.o_proj(output) return output class Glm5NextTextIndexer(nn.Module): """ DeepSeek Sparse Attention (DSA) indexer with k-pool compression for GLM-5.3-Flash. The indexer uses lightweight projections (`wq_b`, `wk`) separate from the main MLA attention path. It scores compressed k-pool candidates, expands selected pools back into raw cache indices, and optionally appends the current incomplete tail pool. **Cache strategy**: the indexer state cache lives on the per-layer `DynamicIndexedLayer` (or the `StaticIndexedLayer` for static caches) inside the shared cache, accessed via `past_key_values.update_indexer()`. """ def __init__(self, config, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.hidden_size: int = config.hidden_size self.n_heads: int = config.index_n_heads self.head_dim: int = config.index_head_dim self.qk_rope_head_dim: int = config.qk_rope_head_dim self.index_topk: int = config.index_topk self.q_lora_rank: int = config.q_lora_rank self.wq_b = nn.Linear(self.q_lora_rank, self.n_heads * self.head_dim, bias=False) self.wk = nn.Linear(self.hidden_size, self.head_dim, bias=False) self.k_norm = nn.LayerNorm(self.head_dim, eps=1e-6) self.weights_proj = nn.Linear(self.hidden_size, self.n_heads, bias=False) self.softmax_scale = self.head_dim**-0.5 self.index_kpool = config.index_kpool self.index_kpool_always_select_tail = config.index_kpool_always_select_tail self.index_kpool_compress_ape = nn.Parameter(torch.zeros(self.index_kpool, self.head_dim)) self.index_kpool_compress_gate = nn.Parameter(torch.zeros(self.head_dim, self.hidden_size)) @torch.no_grad() def forward( self, hidden_states: torch.Tensor, q_resid: torch.Tensor, attention_mask: torch.BoolTensor, past_key_values: Cache | None, ) -> torch.LongTensor: """ Selects the top-k tokens per query for DeepSeek Sparse Attention (DSA) based on grouping pools (and tails). Args: hidden_states: Input hidden states `[B, S, hidden_size]`. q_resid: Query residual from `q_a_layernorm(q_a_proj(x))`, shape `[B, S, q_lora_rank]`. attention_mask: Local boolean padding mask of shape `[B, S]`. past_key_values: Cache object containing the indexer state cache for this layer. Returns: `torch.Tensor`: the `int32` top-k token indices of shape `[B, S, topk]` (or `[B, S, 2*topk - 1]` with tail). The eager / SDPA paths turn these into an additive sparse mask. """ batch_size, seq_len = hidden_states.shape[:2] hidden_shape = (batch_size, seq_len, -1, self.head_dim) q = self.wq_b(q_resid).view(hidden_shape) k = self.k_norm(self.wk(hidden_states)).view(hidden_shape).squeeze(2) gate_scores = F.linear(hidden_states, self.index_kpool_compress_gate) valid_channel = attention_mask.to(k.dtype)[..., None] packed_states = torch.cat([k, gate_scores, valid_channel], dim=-1) kv_len = seq_len current_length = seq_len if past_key_values is not None: cache_layer = past_key_values.layers[self.layer_idx] packed_states = past_key_values.update_indexer(packed_states, self.layer_idx) # Only different on static caches where key is a static full tensor to max len kv_len = cache_layer.keys.shape[-2] current_length = cache_layer.get_seq_length() # Get pools based on the valid key entries (based on padding / causality) valid_keys = packed_states[..., -1].bool() visible_tokens = self.get_visible_tokens( valid_keys=valid_keys, q_length=seq_len, current_length=current_length, ) # Key difference: Score across pools, not on a per token basis pool_keys, pool_indices, pool_valid = self.get_pooled_states(packed_states=packed_states) scores = torch.matmul(q.float(), pool_keys.transpose(-1, -2).float().unsqueeze(1)) scores = F.relu(scores * self.softmax_scale) # Weight per head and sum across heads: [B, S, 1, H] @ [B, S, H, P] -> [B, S, P] weights = self.weights_proj(hidden_states.to(self.weights_proj.weight.dtype)).float() * (self.n_heads**-0.5) index_scores = torch.matmul(weights.unsqueeze(-2), scores).squeeze(-2) # Clamp invalid / static pool ends pool_end = pool_indices[..., -1].clamp(0, kv_len - 1) pool_visible = visible_tokens.gather( dim=-1, index=pool_end[:, None, :].expand(batch_size, seq_len, -1), ) # A pool is selectable only if its final token is visible to the query valid_candidates = pool_visible & pool_valid[:, None] index_scores = index_scores.masked_fill( ~valid_candidates, torch.finfo(index_scores.dtype).min, ) # Similar budgeting as in original but compressed by its pool size select_k = min(self.index_topk // self.index_kpool, index_scores.shape[-1]) # Selection is based on 2 steps # 1. The actual scores (selected indices) # 2. And removing invalid rows based on padding (selected valid) selected = index_scores.topk(select_k, dim=-1).indices batch_idx = torch.arange(batch_size, device=hidden_states.device)[:, None, None] selected_valid = valid_candidates.gather(-1, selected) selected_indices = pool_indices[batch_idx, selected] # Convert selected pools back into the raw tokens # [B, S, K, P] -> [B, S, K * P] topk_indices = selected_indices.flatten(-2) topk_indices = topk_indices.masked_fill( ~selected_valid[..., None].expand_as(selected_indices).flatten(-2), -1, ) output_width = self.index_topk if self.index_kpool_always_select_tail: topk_indices = self.append_visible_tail(topk_indices, visible_tokens, valid_keys) output_width += self.index_kpool - 1 # expanded tail size maximum # Pad so we fill up with invalid entries instead of gathered selections topk_indices = F.pad(topk_indices, (0, output_width - topk_indices.shape[-1]), value=-1) topk_indices = topk_indices[..., :output_width] topk_indices = topk_indices.masked_fill(~attention_mask[..., None], -1) return topk_indices.to(torch.int32) def get_visible_tokens( self, valid_keys: torch.BoolTensor, q_length: int, current_length: int, ) -> torch.BoolTensor: """ Check whether a token is visible for Q based on - Causality - Padding status - Valid keys which is a (cached) variation of the attention mask """ device = valid_keys.device kv_positions = torch.arange(valid_keys.shape[-1], device=device) q_positions = current_length - q_length + torch.arange(q_length, device=device) causal = kv_positions[None, None, :] <= q_positions[None, :, None] return causal & valid_keys[:, None, :] def get_pooled_states( self, packed_states: torch.Tensor, ) -> tuple[torch.Tensor, torch.LongTensor, torch.BoolTensor]: """ (Re-)Build compressed k-pool candidates from the packed state. Pooling starts at the first real token, not raw slot 0. This is the part that makes: [P, P, A, B, C, D, ...] behave like: [A, B, C, D, ...] for k-pool grouping. """ # All states we need # 1. The actual keys # 2. The compressed gate scores # 3. The valid keys based on padding and causality keys, gate_scores, valid_keys = torch.split( packed_states, [self.head_dim, self.head_dim, 1], dim=-1, ) valid_keys = valid_keys.bool().squeeze(-1) # Metadata batch_size, seq_len = keys.shape[:2] number_of_pools = (seq_len + self.index_kpool - 1) // self.index_kpool device = keys.device # Example, index_kpool=4: # [P, P, A, B, C, D, E, F] # ^ # first_key # # pool 0 -> first_key + [0, 1, 2, 3] -> [A, B, C, D] # pool 1 -> first_key + [4, 5, 6, 7] -> [E, F, out, out] first_key = torch.where( valid_keys.any(-1), valid_keys.long().argmax(-1), torch.full((batch_size,), seq_len, dtype=torch.long, device=device), ) pool_offsets = torch.arange(number_of_pools * self.index_kpool, device=device) pool_offsets = pool_offsets.view(1, number_of_pools, self.index_kpool) pool_indices = first_key[:, None, None] + pool_offsets batch_idx = torch.arange(batch_size, device=device)[:, None, None] safe_indices = pool_indices.clamp(0, seq_len - 1) grouped_keys = keys[batch_idx, safe_indices] grouped_gate_scores = gate_scores[batch_idx, safe_indices] grouped_valid_keys = valid_keys[batch_idx, safe_indices] # Only allow those within range (clamp) grouped_valid_keys = grouped_valid_keys & (pool_indices < seq_len) pool_valid = grouped_valid_keys.all(-1) pool_indices = pool_indices.masked_fill(~grouped_valid_keys, -1) # Learn a weighted average over the tokens inside each complete pool logits = grouped_gate_scores.float() + self.index_kpool_compress_ape.float()[None, None] logits = logits.masked_fill(~grouped_valid_keys[..., None], float("-inf")) probabilities = torch.nan_to_num(logits.softmax(dim=2)).to( grouped_keys.dtype ) # nan to num for full invalid pools pool_keys = (probabilities * grouped_keys).sum(dim=2) # Avoids static cache allocated positions keep = pool_valid.any(0) return pool_keys[:, keep], pool_indices[:, keep], pool_valid[:, keep] def append_visible_tail( self, topk_indices: torch.Tensor, token_visible: torch.BoolTensor, key_valid: torch.BoolTensor, ) -> torch.Tensor: """ Append the current incomplete pool as raw token indices. So we if we have a pool size of 4: - [P P A B C D E F] - [A B C D] (selected pool) - [E F] (appended tail) """ if (max_tail_width := self.index_kpool - 1) == 0: return topk_indices batch_size, _, kv_length = token_visible.shape device = token_visible.device # Example, index_kpool=4: # visible keys: [A, B, C, D, E, F] # full pools: [A, B, C, D] # tail: [E, F] # # visible_count = 6 # tail_count = 6 % 4 = 2 # tail_start = first_key + visible_count - tail_count # tail_indices = tail_start + [0, 1, ..., index_kpool - 2] first_key = torch.where( key_valid.any(-1), key_valid.long().argmax(-1), torch.full((batch_size,), kv_length, dtype=torch.long, device=device), ) visible_count = token_visible.long().sum(-1) tail_count = visible_count.remainder(self.index_kpool) tail_offsets = torch.arange(max_tail_width, device=device) tail_start = first_key[:, None] + visible_count - tail_count tail_indices = tail_start[..., None] + tail_offsets # We exclude tails that are just use to fill in positions + those that go beyond the max length tail_valid = (tail_offsets[None, None, :] < tail_count[..., None]) & tail_indices.lt(kv_length) # Also check for padding based tokens kv_idx = tail_indices.clamp(0, kv_length - 1) tail_visible = token_visible.gather(dim=-1, index=kv_idx) # Get the valid conclusion tail_indices = tail_indices.masked_fill(~(tail_valid & tail_visible), -1) return torch.cat([topk_indices, tail_indices], dim=-1) def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor | None, scaling: float, dropout: float = 0.0, **kwargs: Unpack[TransformersKwargs], ): key_states = repeat_kv(key, module.num_key_value_groups) value_states = repeat_kv(value, module.num_key_value_groups) attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value_states) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights class Glm5NextTextAttention(nn.Module): """ DeepSeek-V3.2 DSA extended with **cross-layer top-k sharing**. `config.indexer_types[layer_idx]` decides whether this layer runs its own indexer (`"full"`) or reuses the previous full layer's top-k selection (`"shared"`). `next_skip_topk` signals that the *next* layer will reuse this layer's top-k, so it is propagated upward via `prev_topk_indices`. """ def __init__(self, config: Glm5NextTextConfig, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.attention_dropout = config.attention_dropout self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.q_lora_rank = config.q_lora_rank self.qk_rope_head_dim = config.qk_rope_head_dim self.kv_lora_rank = config.kv_lora_rank self.v_head_dim = config.v_head_dim self.qk_nope_head_dim = config.qk_nope_head_dim self.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads self.is_causal = True self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=config.attention_bias) self.q_a_layernorm = Glm5NextTextRMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False) self.kv_a_proj_with_mqa = nn.Linear( self.hidden_size, config.kv_lora_rank + config.qk_rope_head_dim, bias=config.attention_bias, ) self.kv_a_layernorm = Glm5NextTextRMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) self.kv_b_proj = nn.Linear( config.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), bias=False, ) self.o_proj = nn.Linear( self.num_heads * self.v_head_dim, self.hidden_size, bias=config.attention_bias, ) self.scaling = self.qk_head_dim ** (-0.5) # Refer: https://arxiv.org/abs/2603.12201 for more details. self.skip_topk = config.indexer_types[layer_idx] == "shared" self.indexer = None if self.skip_topk else Glm5NextTextIndexer(config, layer_idx) self.next_skip_topk = ( not self.skip_topk and config.indexer_types[min(layer_idx + 1, len(config.indexer_types) - 1)] == "shared" ) def expand_kv(self, kv_nope: torch.Tensor, k_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: """Expands the compressed latents into key and value states. Args: - kv_nope: key + value without positional encoding, shape [batch_size, 1, seqlen, self.kv_lora_rank] - k_rot: shared key with positional encoding, shape [batch_size, 1, seqlen, self.qk_rope_head_dim] Returns the key and value states, two tensors of shape [batch, num_heads, seq, (k or v)_head_dim]. """ batch_size, _, seq_length, _ = kv_nope.shape key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim) kv_nope = self.kv_b_proj(kv_nope).view(key_shape).transpose(1, 2) k_nope, value_states = torch.split(kv_nope, [self.qk_nope_head_dim, self.v_head_dim], dim=-1) k_rot = k_rot.expand(-1, k_nope.shape[1], -1, -1) # does not affect the underlying storage # Concatenate k_nope and k_rot in a new tensor key_states = kv_nope.new_empty(*kv_nope.shape[:-1], self.qk_nope_head_dim + self.qk_rope_head_dim) key_states[..., : self.qk_nope_head_dim].copy_(k_nope) key_states[..., self.qk_nope_head_dim :].copy_(k_rot) return key_states, value_states def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None, past_key_values: Cache | None = None, prev_topk_indices: torch.Tensor | None = None, **kwargs: Unpack[FlashAttentionKwargs], ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: batch_size, seq_length = hidden_states.shape[:-1] query_shape = (batch_size, seq_length, -1, self.qk_head_dim) # LoRA based path is guaranteed based on the config validation q_resid = self.q_a_layernorm(self.q_a_proj(hidden_states)) query_states = self.q_b_proj(q_resid).view(query_shape).transpose(1, 2) compressed_kv = self.kv_a_proj_with_mqa(hidden_states) kv_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) k_pass = self.kv_a_layernorm(kv_pass).view(batch_size, 1, seq_length, self.kv_lora_rank) k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim) # Cache read / write is performed while latent KV is still compressed if past_key_values is not None: k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx) key_states, value_states = self.expand_kv(k_pass, k_rot) if self.indexer is not None: topk_indices = self.indexer( hidden_states=hidden_states, q_resid=q_resid, attention_mask=attention_mask, past_key_values=past_key_values, ) else: if prev_topk_indices is None: raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.") topk_indices = prev_topk_indices attention_mask = self.build_attention_mask_from_topk( topk_indices=topk_indices, query_states=query_states, kv_length=key_states.shape[2], ) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, **kwargs, ) attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights, topk_indices if self.next_skip_topk else None def build_attention_mask_from_topk( self, topk_indices: torch.Tensor, query_states: torch.Tensor, kv_length: int, ) -> torch.Tensor | None: """ Convert topk_indices into the mask expected by the active backend. Only supporting SDPA and Eager as we have a 3D dependency which cannot be mapped to FA without a custom kernel that can select on a per indices bases per row (query -> topk keys). """ # -1 is invalid as per convention in the indexer # NOTE: The indexer already took care of also excluding padding tokens and causality topk_valid = topk_indices.ge(0) & topk_indices.lt(kv_length) # Clamp only so scatter has a legal index safe_indices = topk_indices.clamp(0, kv_length - 1) selected_counts = torch.zeros( topk_indices.shape[0], # batch size topk_indices.shape[1], # q_length kv_length, # kv_length dtype=torch.int32, device=topk_indices.device, ) selected_counts.scatter_add_(-1, safe_indices, topk_valid.to(torch.int32)) # Final mask 0 == False (not visible), 1 == True (visible) mask = selected_counts.ne(0).unsqueeze(1) # SDPA if self.config._attn_implementation == "sdpa": return mask # Eager min_dtype = torch.finfo(query_states.dtype).min # we need 0s where the tokens should be taken into account, and -inf otherwise (mask is already of boolean type) mask = torch.where(mask, torch.full((), 0.0, device=query_states.device, dtype=query_states.dtype), min_dtype) return mask class Glm5NextTextDecoderLayer(GradientCheckpointingLayer): def __init__(self, config: Glm5NextTextConfig, layer_idx: int): super().__init__() self.block_type = config.layer_types[layer_idx] self.hidden_size = config.hidden_size self.self_attn = ( Glm5NextTextLinearAttention(config, layer_idx) if self.block_type == "linear_attention" else Glm5NextTextAttention(config, layer_idx) ) self.mlp = ( Glm5NextTextMoE(config) if config.mlp_layer_types[layer_idx] == "sparse" else Glm5NextTextMLP(config) ) self.input_layernorm = Glm5NextTextRMSNorm(config.hidden_size, config.rms_norm_eps) self.post_attention_layernorm = Glm5NextTextRMSNorm(config.hidden_size, config.rms_norm_eps) self.attn_hc = Glm5NextTextHyperConnection(config) self.ffn_hc = Glm5NextTextHyperConnection(config) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, use_cache: bool | None = False, position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, prev_topk_indices: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None]: dtype = hidden_states.dtype residual = hidden_states post, comb, hidden_states = self.attn_hc(hidden_states) # Self attn hidden_states = self.input_layernorm(hidden_states) topk_indices = None if self.block_type == "linear_attention": hidden_states = self.self_attn( hidden_states=hidden_states, cache_params=past_key_values, attention_mask=attention_mask, **kwargs, ) else: hidden_states, _, topk_indices = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, use_cache=use_cache, position_embeddings=position_embeddings, prev_topk_indices=prev_topk_indices, **kwargs, ) hidden_states = post.to(dtype).unsqueeze(-1) * hidden_states.unsqueeze(-2) + torch.matmul( comb.to(dtype).transpose(-1, -2), residual ) residual = hidden_states post, comb, hidden_states = self.ffn_hc(hidden_states) # Feed forward hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = post.to(dtype).unsqueeze(-1) * hidden_states.unsqueeze(-2) + torch.matmul( comb.to(dtype).transpose(-1, -2), residual ) return hidden_states, topk_indices @auto_docstring class Glm5NextPreTrainedModel(PreTrainedModel): config: Glm5NextConfig base_model_prefix = "model" supports_gradient_checkpointing = True # needs index based kernel _supports_flash_attn = False _supports_sdpa = True # needs per layer creation, too expensive _supports_flex_attn = False _supports_attention_backend = True _no_split_modules = ["Glm5NextTextDecoderLayer", "Glm5NextVisionBlock", "Glm5NextVisionPatchMerger"] _skip_keys_device_placement = ["past_key_values"] # TODO: this can be fixed but is limited by # 1. assuming the cache name # 2. linear attention not being considered atm _is_stateful = True _can_compile_fullgraph = True _can_record_outputs = { "attentions": Glm5NextTextAttention, "hidden_states": Glm5NextTextDecoderLayer, "router_logits": OutputRecorder(Glm5NextTextTopkRouter, index=0), } _keep_in_fp32_modules_strict = ["e_score_correction_bias", "conv1d", "dt_bias", "A_log"] _keys_to_ignore_on_load_unexpected = [r"layers\.45\.", r"layers\.\d+\.shared_head\."] @torch.no_grad() def _init_weights(self, module): super()._init_weights(module) if isinstance(module, Glm5NextTextForgetGate): # Following FLA initialization # NOTE: This is incredibly important so keep it this way at all costs if module.safe_gate_lower_bound is not None: init.zeros_(module.A_log) else: init.copy_( module.A_log, init.uniform_(module.A_log, a=1.0, b=16.0).log(), ) init.uniform_( module.dt_bias, a=math.log(1e-3), b=math.log(1e-1), ) dt = module.dt_bias.exp().clamp_min(1e-4) # (stable) inverse softplus init.copy_( module.dt_bias, dt + torch.log(-torch.expm1(-dt)), ) elif isinstance(module, Glm5NextTextRMSNormGated): init.ones_(module.weight) elif isinstance(module, Glm5NextTextHyperConnection): init.normal_(module.fn, mean=0.0, std=0.02) init.zeros_(module.base) init.ones_(module.scale) elif isinstance(module, Glm5NextTextExperts): init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range) init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range) elif isinstance(module, Glm5NextTextTopkRouter): init.zeros_(module.e_score_correction_bias) init.normal_(module.weight, mean=0.0, std=self.config.initializer_range) elif isinstance(module, Glm5NextTextIndexer): init.zeros_(module.index_kpool_compress_ape) init.ones_(module.index_kpool_compress_gate) # Do not inherit from DSv4 as it messes modular prefixes up for the PreTrainedModel @auto_docstring class Glm5NextTextModel(Glm5NextPreTrainedModel): config: Glm5NextTextConfig def __init__(self, config): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) self.layers = nn.ModuleList( [Glm5NextTextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = Glm5NextTextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.hc_head = Glm5NextTextHyperHead() # Initialize weights and apply final processing self.post_init() @merge_with_config_defaults @capture_outputs @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, use_cache: bool | None = None, **kwargs: Unpack[TransformersKwargs], ) -> MoeModelOutputWithPast: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if use_cache and past_key_values is None: past_key_values = DynamicCache(config=self.config) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if position_ids is None: past_seen = past_key_values.get_seq_length() if past_key_values is not None else 0 position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen position_ids = position_ids.unsqueeze(0) if not isinstance(causal_mask_mapping := attention_mask, dict): attention_mask = create_recurrent_attention_mask( config=self.config, inputs_embeds=inputs_embeds, attention_mask=attention_mask, past_key_values=past_key_values, ) # Guarantee the mask to exist for the indexer if attention_mask is None: attention_mask = torch.ones( inputs_embeds.shape[0], inputs_embeds.shape[1], dtype=torch.bool, device=inputs_embeds.device, ) attention_mask = attention_mask.bool() causal_mask_mapping = { "indexed_attention": attention_mask, "linear_attention": attention_mask, } hidden_states = inputs_embeds.unsqueeze(2).expand(-1, -1, self.config.hc_mult, -1).contiguous() topk_indices = None for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]): hidden_states, topk_indices = decoder_layer( hidden_states, attention_mask=causal_mask_mapping[self.config.layer_types[i]], position_ids=position_ids, # Key change using NoPE position_embeddings=None, input_ids=input_ids, past_key_values=past_key_values, prev_topk_indices=topk_indices, **kwargs, ) hidden_states = self.norm(self.hc_head(hidden_states)) return MoeModelOutputWithPast(last_hidden_state=hidden_states, past_key_values=past_key_values) class Glm5NextVisionMLP(nn.Module): def __init__(self, config, bias: bool = True): super().__init__() self.hidden_size = config.hidden_size self.intermediate_size = config.intermediate_size self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias) self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=bias) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=bias) self.act_fn = ACT2FN[config.hidden_act] self.swiglu_limit = config.swiglu_limit def forward(self, hidden_state): gate = self.gate_proj(hidden_state) up = self.up_proj(hidden_state) # Key difference using clamping gate = gate.clamp(min=None, max=self.swiglu_limit) up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit) return self.down_proj(self.act_fn(gate) * up) class Glm5NextVisionPatchMerger(nn.Module): def __init__(self, dim: int, context_dim: int, hidden_act: str, swiglu_limit: float, bias: bool = False) -> None: super().__init__() self.proj = nn.Linear(dim, dim, bias=bias) self.post_projection_norm = LayerNorm(dim) self.gate_proj = nn.Linear(dim, context_dim, bias=bias) self.up_proj = nn.Linear(dim, context_dim, bias=bias) self.down_proj = nn.Linear(context_dim, dim, bias=bias) self.act1 = nn.GELU() self.act_fn = ACT2FN[hidden_act] self.swiglu_limit = swiglu_limit def forward(self, hidden_state: torch.Tensor) -> torch.Tensor: hidden_state = self.proj(hidden_state) hidden_state = self.act1(self.post_projection_norm(hidden_state)) gate = self.gate_proj(hidden_state) up = self.up_proj(hidden_state) # Key difference using clamping gate = gate.clamp(min=None, max=self.swiglu_limit) up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit) return self.down_proj(self.act_fn(gate) * up) @use_kernel_forward_from_hub("RMSNorm") class Glm5NextRMSNorm(nn.Module): def __init__(self, hidden_size, eps: float = 1e-6) -> None: """ Glm5NextRMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" def rotate_half(x): """Rotates half the hidden dims of the input.""" x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) def apply_rotary_pos_emb_vision( q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor ) -> tuple[torch.Tensor, torch.Tensor]: orig_q_dtype = q.dtype orig_k_dtype = k.dtype q, k = q.float(), k.float() cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float() q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) q_embed = q_embed.to(orig_q_dtype) k_embed = k_embed.to(orig_k_dtype) return q_embed, k_embed class Glm5NextVisionAttention(nn.Module): def __init__(self, config: Glm5NextVisionConfig) -> None: super().__init__() self.dim = config.hidden_size self.num_heads = config.num_heads self.head_dim = self.dim // self.num_heads self.num_key_value_groups = 1 # needed for eager attention self.qkv = nn.Linear(config.hidden_size, config.hidden_size * 3, bias=config.attention_bias) self.proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attention_bias) self.scaling = self.head_dim**-0.5 self.config = config self.attention_dropout = config.attention_dropout self.is_causal = False self.q_norm = Glm5NextRMSNorm(self.head_dim, eps=config.rms_norm_eps) self.k_norm = Glm5NextRMSNorm(self.head_dim, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, max_seqlen: int | None = None, **kwargs, ) -> torch.Tensor: seq_length = hidden_states.shape[0] query_states, key_states, value_states = ( self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0) ) query_states = self.q_norm(query_states) key_states = self.k_norm(key_states) cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin) query_states = query_states.transpose(0, 1).unsqueeze(0) key_states = key_states.transpose(0, 1).unsqueeze(0) value_states = value_states.transpose(0, 1).unsqueeze(0) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) if is_flash_attention_requested(self.config): # Flash Attention: Use cu_seqlens for variable length attention max_seqlen = get_max_seqlen(cu_seqlens, self.config, kwargs={"max_seqlen": max_seqlen}) attn_output, _ = attention_interface( self, query_states, key_states, value_states, attention_mask=None, scaling=self.scaling, dropout=0.0 if not self.training else self.attention_dropout, cu_seq_lens_q=cu_seqlens, cu_seq_lens_k=cu_seqlens, max_length_q=max_seqlen, max_length_k=max_seqlen, is_causal=False, **kwargs, ) else: # Other implementations: Process each chunk separately lengths = cu_seqlens[1:] - cu_seqlens[:-1] splits = [ torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states) ] attn_outputs = [ attention_interface( self, q, k, v, attention_mask=None, scaling=self.scaling, dropout=0.0 if not self.training else self.attention_dropout, is_causal=False, **kwargs, )[0] for q, k, v in zip(*splits) ] attn_output = torch.cat(attn_outputs, dim=1) attn_output = attn_output.reshape(seq_length, -1).contiguous() attn_output = self.proj(attn_output) return attn_output class Glm5NextVisionBlock(GradientCheckpointingLayer): def __init__(self, config) -> None: super().__init__() self.norm1 = Glm5NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.norm2 = Glm5NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.attn = Glm5NextVisionAttention(config) self.mlp = Glm5NextVisionMLP(config, bias=config.attention_bias) @auto_docstring def forward( self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None, **kwargs, ) -> torch.Tensor: r""" cu_seqlens (`torch.Tensor`): Cumulative sequence lengths used for packed variable-length attention in Flash Attention kernels. """ hidden_states = hidden_states + self.attn( self.norm1(hidden_states), cu_seqlens=cu_seqlens, position_embeddings=position_embeddings, **kwargs, ) hidden_states = hidden_states + self.mlp(self.norm2(hidden_states)) return hidden_states class Glm5NextVisionPatchEmbed(nn.Module): def __init__(self, config: Glm5NextVisionConfig) -> None: super().__init__() self.patch_size = config.patch_size self.temporal_patch_size = config.temporal_patch_size self.in_channels = config.in_channels self.embed_dim = config.hidden_size kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size] self.proj = nn.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: target_dtype = self.proj.weight.dtype hidden_states = hidden_states.view( -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size ) hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim) return hidden_states class Glm5NextVisionRotaryEmbedding(nn.Module): """ Simple axial 2D rope with same freqs used for H and W grids. The freqs are pre-computed using `head-dim//4` which is later used to concat H and W positions. The final angles rotate over the whole head dim, no partial rotation involved. """ def __init__(self, config: Glm5NextVisionConfig): super().__init__() self.config = config self.rope_type = self.config.rope_parameters["rope_type"] rope_init_fn: Callable = self.compute_axial_rope_parameters if self.rope_type != "axial": raise ValueError(f"{self.__class__.__name__} supports only axial rope, but requested {self.rope_type}") inv_freq, self.attention_scaling = rope_init_fn(self.config) self.inv_freq = nn.Buffer(inv_freq, persistent=False) self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False) @staticmethod def compute_axial_rope_parameters( config: Glm5NextVisionConfig, device=None, **kwargs ) -> tuple[torch.Tensor, float]: """ Computes the inverse frequencies according to the original RoPE implementation Args: config ([`~transformers.PreTrainedConfig`]): The model configuration. Returns: Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE). """ base = config.rope_parameters["rope_theta"] dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads spatial_dim = dim // 2 attention_factor = 1.0 # Unused in this type of RoPE inv_freq = 1.0 / (base ** (torch.arange(0, spatial_dim, 2, dtype=torch.float) / spatial_dim)) return inv_freq.to(device), attention_factor @torch.no_grad() def forward(self, x, position_ids): # position_ids: (2, N) — row 0 = h coords, row 1 = w coords freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float) cos = freqs.cos() * self.attention_scaling sin = freqs.sin() * self.attention_scaling cos = self.recomposition_frequencies(cos) sin = self.recomposition_frequencies(sin) return cos, sin def recomposition_frequencies(self, freq): """ Recompose the frequencies into the final spatial layout used per each grid. """ freq_h, freq_w = freq[:, 0], freq[:, 1] freq_hw = torch.cat([freq_h, freq_w], dim=-1) return torch.cat([freq_hw, freq_hw], dim=-1) @auto_docstring class Glm5NextVisionModel(Glm5NextPreTrainedModel): config: Glm5NextVisionConfig input_modalities = ("image", "video") _no_split_modules = ["Glm5NextVisionBlock"] _can_record_outputs = { "hidden_states": Glm5NextVisionBlock, "attentions": Glm5NextVisionAttention, } def __init__(self, config) -> None: super().__init__(config) self.spatial_merge_size = config.spatial_merge_size self.patch_size = config.patch_size self.patch_embed = Glm5NextVisionPatchEmbed(config) self.rotary_pos_emb = Glm5NextVisionRotaryEmbedding(config) self.blocks = nn.ModuleList([Glm5NextVisionBlock(config) for _ in range(config.depth)]) self.merger = Glm5NextVisionPatchMerger( dim=config.out_hidden_size, context_dim=config.projection_intermediate_size, hidden_act=config.hidden_act, swiglu_limit=config.swiglu_limit, ) self.downsample = nn.Conv2d( in_channels=config.hidden_size, out_channels=config.out_hidden_size, kernel_size=config.spatial_merge_size, stride=config.spatial_merge_size, ) self.post_layernorm = Glm5NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.post_init() @merge_with_config_defaults @capture_outputs @auto_docstring def forward( self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs, ) -> torch.Tensor: r""" hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`): The final hidden states of the model. grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`): The temporal, height and width of feature shape of each image in LLM. Returns: `torch.Tensor`: hidden_states. """ position_ids = get_vision_position_ids(grid_thw, self.spatial_merge_size, kwargs=kwargs) cu_seqlens, max_seqlen = get_vision_attention_seqlens(grid_thw, self.config, kwargs=kwargs) hidden_states = self.patch_embed(hidden_states) position_embeddings = self.rotary_pos_emb(hidden_states, position_ids) for blk in self.blocks: hidden_states = blk( hidden_states, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, position_embeddings=position_embeddings, ) hidden_states = self.post_layernorm(hidden_states) hidden_states = hidden_states.view( -1, self.spatial_merge_size, self.spatial_merge_size, hidden_states.shape[-1] ) hidden_states = hidden_states.permute(0, 3, 1, 2) hidden_states = self.downsample(hidden_states).view(-1, self.config.out_hidden_size) merged_hidden_states = self.merger(hidden_states) return BaseModelOutputWithPooling( last_hidden_state=hidden_states, pooler_output=merged_hidden_states, ) @auto_docstring class Glm5NextModel(Glm5NextPreTrainedModel): base_model_prefix = "model" # Reference: fix gemma3 grad acc #37208 accepts_loss_kwargs = False config: Glm5NextConfig def __init__(self, config): super().__init__(config) self.visual = Glm5NextVisionModel._from_config(config.vision_config) self.language_model = Glm5NextTextModel._from_config(config.text_config) # Initialize weights and apply final processing self.post_init() @accepts_precomputed_kwargs(modality="video") @can_return_tuple @auto_docstring def get_video_features( self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPast: # Same as in `Glm46V` # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames t = video_grid_thw[:, 0] hw = video_grid_thw[:, 1:] # repeat each (h,w) row `t` times flattened_hw = torch.repeat_interleave(hw, t, dim=0) prefix_ones = video_grid_thw.new_ones(flattened_hw.shape[0], 1) flattened_video_grid_thw = torch.cat([prefix_ones, flattened_hw], dim=1) pixel_values_videos = pixel_values_videos.type(self.visual.dtype) vision_outputs = self.visual(pixel_values_videos, grid_thw=flattened_video_grid_thw, **kwargs) split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist() vision_outputs.pooler_output = torch.split(vision_outputs.pooler_output, split_sizes) return vision_outputs @accepts_precomputed_kwargs(modality="image") @can_return_tuple @auto_docstring def get_image_features( self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: pixel_values = pixel_values.type(self.visual.dtype) vision_outputs = self.visual(pixel_values, grid_thw=image_grid_thw, **kwargs) split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist() vision_outputs.pooler_output = torch.split(vision_outputs.pooler_output, split_sizes) return vision_outputs def get_placeholder_mask( self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor | None = None, video_features: torch.FloatTensor | None = None, ): """ Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is equal to the length of multimodal features. If the lengths are different, an error is raised. """ if input_ids is None: special_mm_mask = inputs_embeds == self.get_input_embeddings()( torch.full((), self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device) ) special_mm_mask = special_mm_mask.all(-1) video_start_mask = inputs_embeds == self.get_input_embeddings()( torch.full((), self.config.video_start_token_id, dtype=torch.long, device=inputs_embeds.device) ) video_start_mask = video_start_mask.all(-1) video_end_mask = inputs_embeds == self.get_input_embeddings()( torch.full((), self.config.video_end_token_id, dtype=torch.long, device=inputs_embeds.device) ) video_end_mask = video_end_mask.all(-1) in_video_span = video_start_mask.cumsum(-1) > video_end_mask.cumsum(-1) else: special_mm_mask = input_ids == self.config.image_token_id in_video_span = (input_ids == self.config.video_start_token_id).cumsum(-1) > ( input_ids == self.config.video_end_token_id ).cumsum(-1) # Core difference to other VLMs as img token == vid token so we differentiate by start/end spans instead special_image_mask = special_mm_mask & ~in_video_span special_video_mask = special_mm_mask & in_video_span n_image_tokens = special_image_mask.sum() special_image_mask = special_image_mask.unsqueeze(-1).to(inputs_embeds.device) if image_features is not None: torch_compilable_check( n_image_tokens * inputs_embeds.shape[-1] == image_features.numel(), f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}", ) n_video_tokens = special_video_mask.sum() special_video_mask = special_video_mask.unsqueeze(-1).to(inputs_embeds.device) if video_features is not None: torch_compilable_check( n_video_tokens * inputs_embeds.shape[-1] == video_features.numel(), f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}", ) return special_image_mask, special_video_mask @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, use_cache: bool | None = None, pixel_values: torch.Tensor | None = None, pixel_values_videos: torch.FloatTensor | None = None, image_grid_thw: torch.LongTensor | None = None, video_grid_thw: torch.LongTensor | None = None, mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPast: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if (pixel_values is not None or pixel_values_videos is not None) and mm_encoder_outputs is not None: raise ValueError( "You cannot specify both pixel_values/pixel_values_videos and mm_encoder_outputs at the same time" ) if inputs_embeds is None: inputs_embeds = self.get_input_embeddings()(input_ids) mm_encoder_outputs = mm_encoder_outputs if mm_encoder_outputs is not None else {} if mm_encoder_outputs.get("image") is None and pixel_values is not None: mm_encoder_outputs["image"] = self.get_image_features( pixel_values, image_grid_thw, return_dict=True, **kwargs ) if mm_encoder_outputs.get("video") is None and pixel_values_videos is not None: mm_encoder_outputs["video"] = self.get_video_features( pixel_values_videos, video_grid_thw, return_dict=True, **kwargs ) if mm_encoder_outputs.get("image") is not None: image_embeds = torch.cat(mm_encoder_outputs["image"].pooler_output, dim=0).to( inputs_embeds.device, inputs_embeds.dtype ) image_mask, _ = self.get_placeholder_mask( input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds ) inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds) if mm_encoder_outputs.get("video") is not None: video_embeds = torch.cat(mm_encoder_outputs["video"].pooler_output, dim=0).to( inputs_embeds.device, inputs_embeds.dtype ) _, video_mask = self.get_placeholder_mask( input_ids, inputs_embeds=inputs_embeds, video_features=video_embeds ) inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds) outputs = self.language_model( input_ids=None, position_ids=position_ids, attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, **kwargs, ) # Only change is the output type to Moe return MoeModelOutputWithPast( last_hidden_state=outputs.last_hidden_state, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, router_logits=outputs.router_logits, ) def load_balancing_loss_func( gate_logits: torch.Tensor | tuple[torch.Tensor] | None, num_experts: int | None = None, top_k=2, attention_mask: torch.Tensor | None = None, ) -> torch.Tensor | int: r""" Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between experts is too unbalanced. Args: gate_logits: Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of shape [batch_size X sequence_length, num_experts]. num_experts: Number of experts top_k: The number of experts to route per-token, can be also interpreted as the `top-k` routing parameter. attention_mask (`torch.Tensor`, *optional*): The attention_mask used in forward function shape [batch_size X sequence_length] if not None. Returns: The auxiliary loss. """ if gate_logits is None or not isinstance(gate_logits, tuple): return 0 # Accumulate assignment counts and probability sums layer by layer, normalizing at the end, # so peak memory stays O(seq_len * num_experts) regardless of the number of layers. compute_device = gate_logits[0].device tokens_per_expert_sum = torch.zeros(num_experts, dtype=torch.float32, device=compute_device) router_prob_sum = torch.zeros(num_experts, dtype=torch.float32, device=compute_device) total_rows = 0.0 if attention_mask is not None: # The same flat mask applies to every layer's [batch_size * sequence_length] rows. flat_mask = attention_mask.reshape(-1).to(device=compute_device, dtype=torch.float32) for layer_gate in gate_logits: routing_weights = torch.nn.functional.softmax(layer_gate.to(compute_device), dim=-1) _, selected_experts = torch.topk(routing_weights, top_k, dim=-1) if attention_mask is None: # Count of top-k assignments per expert tokens_per_expert_sum = ( tokens_per_expert_sum + torch.bincount(selected_experts.reshape(-1), minlength=num_experts).float() ) # Sum of routing probabilities per expert router_prob_sum = router_prob_sum + routing_weights.float().sum(dim=0) total_rows = total_rows + routing_weights.shape[0] else: # Same reductions, weighted by the attention mask to exclude padding tokens tokens_per_expert_sum = tokens_per_expert_sum + torch.zeros( num_experts, dtype=torch.float32, device=compute_device ).scatter_add_(0, selected_experts.reshape(-1), flat_mask.repeat_interleave(top_k)) router_prob_sum = router_prob_sum + (routing_weights.float() * flat_mask.unsqueeze(-1)).sum(dim=0) total_rows = total_rows + flat_mask.sum() tokens_per_expert = tokens_per_expert_sum / total_rows router_prob_per_expert = router_prob_sum / total_rows overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0)) return overall_loss * num_experts @auto_docstring class Glm5NextForConditionalGeneration(Glm5NextPreTrainedModel, GenerationMixin): """ Main Glm5Next conditional generation class. """ _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"} _tp_plan = {"lm_head": "colwise_gather_output"} # Reference: fix gemma3 grad acc #37208 accepts_loss_kwargs = False def __init__(self, config): super().__init__(config) self.model = Glm5NextModel(config) self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False) self.router_aux_loss_coef = config.text_config.router_aux_loss_coef self.num_experts = config.text_config.num_local_experts self.num_experts_per_tok = config.text_config.num_experts_per_tok self.post_init() @auto_docstring def get_video_features( self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: r""" pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`): The tensors corresponding to the input videos. """ return self.model.get_video_features(pixel_values_videos, video_grid_thw, **kwargs) @auto_docstring def get_image_features( self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPooling: r""" pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`): The tensors corresponding to the input images. """ return self.model.get_image_features(pixel_values, image_grid_thw, **kwargs) @can_return_tuple @auto_docstring def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, labels: torch.LongTensor | None = None, use_cache: bool | None = None, pixel_values: torch.Tensor | None = None, pixel_values_videos: torch.FloatTensor | None = None, image_grid_thw: torch.LongTensor | None = None, video_grid_thw: torch.LongTensor | None = None, output_router_logits: bool | None = None, mm_token_type_ids: torch.IntTensor | None = None, logits_to_keep: int | torch.Tensor = 0, mm_encoder_outputs: dict[str, BaseModelOutputWithPooling] | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple | MoeCausalLMOutputWithPast: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Example: ```python >>> from transformers import AutoProcessor, Glm5NextForConditionalGeneration >>> import torch >>> model = Glm5NextForConditionalGeneration.from_pretrained("zai-org/GLM-5.3-Flash") >>> processor = AutoProcessor.from_pretrained("zai-org/GLM-5.3-Flash") >>> messages = [ ... { ... "role": "user", ... "content": [ ... {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"}, ... {"type": "text", "text": "Describe the image."}, ... ], ... } ... ] >>> inputs = processor.apply_chat_template( ... messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt" ... ) >>> inputs = {k: v.to(model.device) if isinstance(v, torch.Tensor) else v for k, v in inputs.items()} >>> generated_ids = model.generate(**inputs, max_new_tokens=64) ``` """ output_router_logits = ( output_router_logits if output_router_logits is not None else self.config.text_config.output_router_logits ) outputs = self.model( input_ids=input_ids, pixel_values=pixel_values, pixel_values_videos=pixel_values_videos, image_grid_thw=image_grid_thw, video_grid_thw=video_grid_thw, position_ids=position_ids, attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_router_logits=output_router_logits, mm_encoder_outputs=mm_encoder_outputs, **kwargs, ) hidden_states = outputs.last_hidden_state slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: loss = self.loss_function( logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs ) aux_loss = None if output_router_logits: aux_loss = load_balancing_loss_func( outputs.router_logits, self.num_experts, self.num_experts_per_tok, attention_mask, ) if labels is not None: loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device return MoeCausalLMOutputWithPast( loss=loss, aux_loss=aux_loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, router_logits=outputs.router_logits, ) def _expand_multimodal_outputs( self, input_ids: torch.LongTensor, mm_encoder_output: dict, expand_size: int = 1, inputs_embeds: torch.LongTensor | None = None, ) -> dict[str, dict]: # override -> model uses the same config key for image and video placeholders def repeat_tensor_or_list(inputs: list | torch.Tensor, repeat_times: int): # Tensor of size [bs, seqlen, dim] where `bs` is number of images in this text sample # Each text can have 1+ images associated with it # Inteleaving on first dim does the same thing as `input_ids.repeat_interlave` in leading batch dim! if isinstance(inputs, torch.Tensor): return inputs.repeat_interleave(repeat_times, dim=0) else: # List of `bs` length where each entry is a tensor (seqlen, dim) is also repeat interleaved return [beam_entry for entry in inputs for beam_entry in [entry] * repeat_times] for modality in ["image", "video"]: modalily_outputs = mm_encoder_output.get(modality) if modalily_outputs is None or getattr(modalily_outputs, "pooler_output", None) is None: continue # 1. compute cumulative number of placehlder tokens per sample and per each encoded mm-data token_id_key = f"{modality}_token_id" if (input_ids is None or input_ids.numel() == 0) and inputs_embeds is not None: special_image_mask = inputs_embeds == self.get_input_embeddings()( torch.full((), self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device) ) num_image_tokens_in_text = special_image_mask.all(-1).sum(-1) else: num_image_tokens_in_text = (input_ids == getattr(self.config, token_id_key)).sum(-1) num_image_tokens_in_vision = [len(out) for out in modalily_outputs.pooler_output] num_image_tokens_in_text = list(accumulate(num_image_tokens_in_text)) num_image_tokens_in_vision = list(accumulate(num_image_tokens_in_vision)) # 2. Find offsets to split encoder output into separate groups per text. In a single batch # we might get a text with single image and another with two images, so the most reliable # way to split dynamic-sized images is by checking number of placeholders and encoder output lengths! offsets = [0] + [ i + 1 for i, num in enumerate(num_image_tokens_in_vision) if num in num_image_tokens_in_text ] is_split_images = not isinstance(modalily_outputs.pooler_output, torch.Tensor) modalily_outputs.pooler_output = [ out for start, end in zip(offsets[:-1], offsets[1:]) for out in repeat_tensor_or_list(modalily_outputs.pooler_output[start:end], expand_size) ] # 3. if `pooler_output` was a tensor, cat it back to follow model expectations if not is_split_images: modalily_outputs.pooler_output = torch.stack(modalily_outputs.pooler_output, dim=0) return mm_encoder_output @staticmethod def create_masks_for_generate(config, inputs_embeds, attention_mask, past_key_values, **_): # We only use the base 2D mask as the indexer is reliant on the padding, not the expanded masks. # I.e. 4D masks are built afterwards after subsets have been selected in the indexer. # Linear attention can reuse the mask as is, making the layer type difference only # necessary for the cache. attention_mask = create_recurrent_attention_mask( config=config.get_text_config(), inputs_embeds=inputs_embeds, attention_mask=attention_mask, past_key_values=past_key_values, ) # Guarantee the mask to exist for the indexer if attention_mask is None: attention_mask = torch.ones( inputs_embeds.shape[0], inputs_embeds.shape[1], dtype=torch.bool, device=inputs_embeds.device, ) attention_mask = attention_mask.bool() return {"indexed_attention": attention_mask, "linear_attention": attention_mask} __all__ = [ "Glm5NextPreTrainedModel", "Glm5NextTextModel", "Glm5NextVisionModel", "Glm5NextModel", "Glm5NextForConditionalGeneration", ]