Ripple-65B-Preview-tf5-patch / modeling_ricursive.py
zx-modelcloud's picture
Simplify Ripple patch documentation and remove internal source references
f052c8d verified
Raw History Blame
51.2 kB
"""HuggingFace modeling implementation for the Ricursive transformer architecture. Designed for inference with vLLM (via --trust-remote-code). Key differences from standard Llama/Mistral: - PeriNorm blocks: 4 layer norms per block (pre/post attention + pre/post FFN) - Gated attention: elementwise sigmoid gate on attention output - Mixed local/global layers: local = sliding window + RoPE, global = full attention + no RoPE - Per-head Q/K RMSNorm Transformers 5.x compatibility (2026-11-05) ------------------------------------------ `generate()` used to break on the very first prefill because `prepare_inputs_for_generation` sliced `input_ids` with a `None` `cache_position`, and the causal/sliding-window mask builders assumed the transformers 4.x cache API. The fixes are applied inline below: * `prepare_inputs_for_generation` rebuilds a missing `cache_position`, slices only when the cache is warm, derives per-row `position_ids` from the padding mask, and moves inputs to the model device. * `_update_causal_mask` sizes the key axis from `cache_position` instead of the removed `get_max_cache_shape()`, emits `0.0` for visible keys, and never returns an all-masked single-token decode row. * `_apply_sliding_window_mask` anchors key positions at the first key the layer actually holds (the cache evicts from the left) and is skipped when the window cannot bind. * `RicursiveForCausalLM` inherits `GenerationMixin` explicitly; transformers 5 no longer injects it into remote-code classes. DFlash2 inference fixes (2026-11-06) ----------------------------------- * `RicursiveAttention.forward` builds an end-anchored causal mask for a multi-token verify block running against a warm cache. SDPA's implicit mask is top-left aligned, so with `q_len < kv_len` it hid valid context. Preallocated caches (`StaticCache`) return their whole capacity from `update()`, so the mask is now built from the layer's *actual* cache length and query positions; the unused tail is trimmed before attention. * `RicursiveForCausalLM.forward` honours `logits_to_keep` by slicing the LM head *input*, so the full `[batch, seq, vocab]` projection is never materialised; `prepare_inputs_for_generation` requests one row. Labeled calls keep the full sequence so the shifted causal-LM loss stays defined, then trim only the returned logits to the requested width. * `RicursiveMLP.forward` chunks the row dimension (`RIPPLE_MLP_CHUNK`, default 2048, `0` disables) to bound the prefill activation peak. Every op is row-wise, so the result is exact. Cache-alignment fixes from review (2026-11-06) ---------------------------------------------- * A mask longer than the key tensor is now aligned to the absolute positions the layer actually holds, `[total_kv - kv_len, total_kv)`, instead of to the tail of the mask. With a mask sized to the allocated cache *capacity* (`StaticCache`), the old `[-kv_len:]` slice selected unwritten slots after sliding-window eviction and silently returned wrong attention. * `RicursiveAttention.forward` forwards `cache_position` to `past_key_value.update()`. A transformers 4.x `StaticCache` otherwise falls back to `arange(q_len)` and the first decode step overwrites the prompt instead of appending after it. Model contract fixes (2026-11-09) ------------------------------- * Generation supports embeddings-only prompts, continuation buffers, explicit position advancement, short masks, and requested diagnostic outputs. * Attention preserves caller 4D masks, reports requested weights without changing SDPA outputs, and trims unused static capacity before expanding KV heads. * Default caches follow the hybrid layer configuration on Transformers 5; Transformers 4 keeps a plain, crop-compatible full cache. Training implements gradient checkpointing and computes causal-LM loss in FP32. * Loss honors accumulation-wide token counts; TP reshapes use local head counts. * Tied embeddings support Transformers 4/5, and window masks preserve exclusions. """ import os from functools import partial from typing import Optional, Tuple, Union import torch import torch.nn as nn import torch.nn.functional as F from transformers import PreTrainedModel from transformers.cache_utils import Cache, DynamicCache from transformers.modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, ) try: # transformers >= 4.26, re-exported from here in 5.x from transformers.generation.utils import GenerationMixin except ImportError: # pragma: no cover - very old transformers from transformers.generation import GenerationMixin from .configuration_ricursive import RicursiveConfig # Set when the transformers-5.x `generate()` fixes above are present natively. # Downstream loaders that monkeypatch older checkpoints check this marker and # stand down instead of re-installing the equivalent runtime patch. RICURSIVE_TF5_COMPAT = "2026-11-09.1" def _transformers_major_version() -> int: """Return the installed Transformers major version, or 0 when unknown.""" try: from transformers import __version__ as version except ImportError: # pragma: no cover - Transformers is a hard dependency return 0 try: return int(version.split(".", 1)[0]) except (AttributeError, ValueError): return 0 def _new_default_cache(config: RicursiveConfig) -> DynamicCache: """Create the default KV cache for a forward pass. Transformers 5 builds a hybrid cache from ``config``: sliding layers keep only their window while full layers grow, which bounds long-context decode memory. Transformers 4 also accepts ``config`` but its ``DynamicSlidingWindowLayer`` cannot be cropped after it has seen more tokens than the window, so assisted decoding and any cache rollback raise ``ValueError: Cannot crop a DynamicSlidingWindowLayer``. Keep the plain full-attention cache on 4.x, matching the pre-config default. """ if _transformers_major_version() >= 5: try: return DynamicCache(config=config) except TypeError: # A 5.x build without configured hybrid caches. pass return DynamicCache() class RicursiveRMSNorm(nn.Module): def __init__(self, hidden_size: int, eps: float = 1e-6): super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.eps = eps def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: input_dtype = hidden_states.dtype hidden_states = hidden_states.float() variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.eps) return self.weight * hidden_states.to(input_dtype) class RicursiveRotaryEmbedding(nn.Module): def __init__(self, dim: int, theta: float = 10000.0): super().__init__() self.dim = dim self.theta = theta # NOTE: inv_freq is computed on-the-fly in forward() rather than stored # as a buffer, because non-persistent buffers get corrupted during # from_pretrained() loading. @torch.no_grad() def forward(self, x: torch.Tensor, position_ids: torch.Tensor): inv_freq = 1.0 / ( self.theta ** (torch.arange(0, self.dim, 2, device=x.device).float() / self.dim) ) inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() sin = emb.sin() return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) def rotate_half(x: torch.Tensor) -> torch.Tensor: x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) def apply_rotary_pos_emb( q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, ) -> Tuple[torch.Tensor, torch.Tensor]: cos = cos.unsqueeze(1) # (B, 1, T, dim) sin = sin.unsqueeze(1) # (B, 1, T, dim) # OLMo-core applies RoPE in full precision (float32) for numerical stability orig_dtype = q.dtype q_float, k_float = q.float(), k.float() cos_float, sin_float = cos.float(), sin.float() q_embed = (q_float * cos_float) + (rotate_half(q_float) * sin_float) k_embed = (k_float * cos_float) + (rotate_half(k_float) * sin_float) return q_embed.to(orig_dtype), k_embed.to(orig_dtype) def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: if n_rep == 1: return hidden_states batch, num_kv_heads, slen, head_dim = hidden_states.shape hidden_states = hidden_states[:, :, None, :, :].expand( batch, num_kv_heads, n_rep, slen, head_dim ) return hidden_states.reshape(batch, num_kv_heads * n_rep, slen, head_dim) class RicursiveAttention(nn.Module): def __init__(self, config: RicursiveConfig, layer_idx: int): super().__init__() self.config = config self.layer_idx = layer_idx self.layer_type = config.layer_types[layer_idx] self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.num_kv_heads = config.num_key_value_heads self.head_dim = config.head_dim self.num_kv_groups = self.num_heads // self.num_kv_heads self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) if config.use_valres: self.value_res_lambda = nn.Parameter(torch.zeros(1)) # Per-head Q/K norms (RMSNorm on head_dim) self.q_norm = RicursiveRMSNorm(self.head_dim, eps=config.rms_norm_eps) self.k_norm = RicursiveRMSNorm(self.head_dim, eps=config.rms_norm_eps) # Gated attention if config.use_gated_attention: self.g_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) # RoPE only for local (sliding_attention) layers if self.layer_type == "sliding_attention": self.rotary_emb = RicursiveRotaryEmbedding(self.head_dim, theta=config.rope_theta) def _apply_sliding_window_mask( self, attention_mask: Optional[torch.Tensor], q_len: int, kv_len: int, dtype: torch.dtype, device: torch.device, cache_position: Optional[torch.LongTensor], total_kv: Optional[int] = None, ) -> Optional[torch.Tensor]: if self.layer_type != "sliding_attention": return attention_mask # transformers >= 5 builds a hybrid cache: a sliding layer holds at most # `sliding_window` keys while a full-attention layer keeps growing, so the # shared mask produced by `_update_causal_mask` can be longer than this # layer's key tensor. Keep only the keys this layer actually holds. if attention_mask is not None and attention_mask.shape[-1] > kv_len: mask_len = attention_mask.shape[-1] if total_kv is not None: # The retained keys sit at `[total_kv - kv_len, total_kv)`. start = max(0, min(total_kv - kv_len, mask_len - kv_len)) attention_mask = attention_mask[..., start : start + kv_len] else: attention_mask = attention_mask[..., -kv_len:] if cache_position is not None: query_positions = cache_position.to(device=device).reshape(-1) else: query_positions = torch.arange(kv_len - q_len, kv_len, device=device) window = self.config.sliding_window if window is None: return attention_mask if kv_len <= window: # The window cannot exclude any key, so the result is either the # unchanged (already causal) mask or, without one, plain causal # attention -- which SDPA applies through `is_causal`. if attention_mask is None: # SDPA's implicit causal mask is top-left aligned. It only # matches this query block when it is the whole sequence # (`q_len == kv_len`) or a single decode position # (`q_len == 1`). A multi-token block verified against a warm # cache sits at the *end* of the key span, so fall through and # build the end-anchored mask below. if q_len == 1 or q_len == kv_len: return None if q_len == 1: return attention_mask # Keys are contiguous and end at the newest query position. After the cache # evicts from the left they no longer start at position 0, so anchor the # range at the first key the layer holds instead of at zero. key_positions = torch.arange(kv_len, device=device) + ( query_positions[-1] - kv_len + 1 ) allowed = ( (key_positions[None, :] <= query_positions[:, None]) & (key_positions[None, :] >= query_positions[:, None] - (window - 1)) )[None, None, :, :] if attention_mask is None: return allowed if attention_mask.dtype == torch.bool: return attention_mask & allowed # Finite minima would reopen an existing -inf exclusion on an empty row. return attention_mask.masked_fill(~allowed, float("-inf")) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, value_residual: Optional[torch.Tensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.shape # Project Q, K, V q = self.q_proj(hidden_states) k = self.k_proj(hidden_states) value_states = hidden_states if value_residual is not None: if not self.config.use_valres: raise RuntimeError("value_residual was supplied but config.use_valres is false") blend = torch.sigmoid(self.value_res_lambda[0]).to(dtype=hidden_states.dtype) value_states = torch.lerp( value_residual.to(dtype=hidden_states.dtype), hidden_states, blend ) v = self.v_proj(value_states) # Infer rank-local head counts from projections after tensor parallel sharding. q = q.view(bsz, q_len, -1, self.head_dim) k = k.view(bsz, q_len, -1, self.head_dim) v = v.view(bsz, q_len, -1, self.head_dim) # Per-head Q/K norm (applied before transpose, on head_dim) q = self.q_norm(q) k = self.k_norm(k) # Transpose to (B, n_heads, T, head_dim) for attention q = q.transpose(1, 2) k = k.transpose(1, 2) v = v.transpose(1, 2) # RoPE only for local layers if self.layer_type == "sliding_attention": cos, sin = self.rotary_emb(v, position_ids) q, k = apply_rotary_pos_emb(q, k, cos, sin) # KV cache total_kv = None if past_key_value is not None: # Forward the absolute positions of the tokens being written. A # transformers 4.x `StaticCache` falls back to `arange(q_len)` when # `cache_position` is absent, so the first decode step would write # at slots 0..q_len-1 and overwrite the prompt instead of appending # after it. `DynamicCache`/5.x caches ignore the extra kwarg. cache_kwargs = ( {"cache_position": cache_position} if cache_position is not None else None ) if cache_kwargs is not None: try: k, v = past_key_value.update(k, v, self.layer_idx, cache_kwargs) except TypeError: # cache without a `cache_kwargs` parameter k, v = past_key_value.update(k, v, self.layer_idx) else: k, v = past_key_value.update(k, v, self.layer_idx) # Number of keys this layer has actually written. A preallocated # (StaticCache) layer returns its full capacity from `update()`, # including slots that have not been filled yet; an evicting # (sliding-window) layer reports more than the trailing window it # returns. The mask must use the real positions, not the tensor size. total_kv = _cache_layer_seq_length(past_key_value, self.layer_idx) # Drop the unwritten tail of a preallocated cache. Without this a # six-token prefill into a sixteen-slot StaticCache looks like a # ten-token prefix, and the end-anchored verify mask below would let # each query attend slots belonging to a different absolute position. if total_kv is not None and 0 < total_kv < k.shape[2]: k = k[:, :, :total_kv] v = v[:, :, :total_kv] # Expand only written slots, not the entire capacity of a static cache. num_kv_groups = q.shape[1] // k.shape[1] k = repeat_kv(k, num_kv_groups) v = repeat_kv(v, num_kv_groups) # Align an incoming mask that is longer than the keys this layer holds. # The shared mask is built over absolute key positions # `0..target_length-1`; keep the slice that matches the keys present. # # The keys returned by `update()` occupy the absolute positions # `[total_kv - kv_len, total_kv)`. Anchoring on the tensor tail # (`[-kv_len:]`) is only correct when the mask length equals # `total_kv`; with a mask sized to the *allocated cache capacity* it # selects the unwritten tail instead of the retained window and # silently produces wrong attention. if attention_mask is not None and attention_mask.shape[-1] > k.shape[2]: kv_len_now = k.shape[2] mask_len = attention_mask.shape[-1] if total_kv is not None: start = max(0, min(total_kv - kv_len_now, mask_len - kv_len_now)) attention_mask = attention_mask[..., start : start + kv_len_now] else: # Unknown cache: the keys are the trailing window. attention_mask = attention_mask[..., -kv_len_now:] attention_mask = self._apply_sliding_window_mask( attention_mask=attention_mask, q_len=q.shape[2], kv_len=k.shape[2], dtype=q.dtype, device=q.device, cache_position=cache_position, total_kv=total_kv, ) if attention_mask is None and q.shape[2] > 1 and k.shape[2] > q.shape[2]: # Multi-token verify block against a warm cache. SDPA's implicit # causal mask is top-left aligned, so when `q_len < kv_len` it hides # every key before the block. Build the mask from the actual query # positions (and the absolute positions of the returned keys) rather # than assuming the key tensor starts at position 0. if cache_position is not None and cache_position.numel() == q.shape[2]: query_positions = cache_position.to( device=q.device, dtype=torch.long ).reshape(-1) else: query_positions = torch.arange( k.shape[2] - q.shape[2], k.shape[2], device=q.device ) if total_kv is not None and total_kv > k.shape[2]: # Evicting cache: the returned keys are the trailing window that # ends at the newest query, not positions 0..kv_len-1. last_query = int(query_positions[-1].item()) key_positions = torch.arange( last_query - k.shape[2] + 1, last_query + 1, device=q.device ) else: key_positions = torch.arange(k.shape[2], device=q.device) attention_mask = ( key_positions[None, :] <= query_positions[:, None] )[None, None] # Attention (use is_causal when no explicit mask for correct causal behavior) # Cast to plain bool to avoid SymBool issues with torch.compile is_causal = bool(attention_mask is None and q.shape[2] > 1) # Keep the SDPA result even when returning diagnostic weights. Rebuilding # the output from rounded BF16 probabilities changes logits and token IDs. attn_output = F.scaled_dot_product_attention( q, k, v, attn_mask=attention_mask, dropout_p=0.0, is_causal=is_causal ) attn_weights = None if output_attentions: # SDPA does not expose weights; compute them separately only on request. weight_keys = k with torch.autocast(device_type=q.device.type, enabled=False): scores = torch.matmul(q.float(), weight_keys.float().transpose(-1, -2)) * self.head_dim**-0.5 if is_causal: causal = torch.ones(q_len, weight_keys.shape[2], dtype=torch.bool, device=q.device).tril() scores = scores.masked_fill(~causal, float("-inf")) if attention_mask is not None: if attention_mask.dtype == torch.bool: scores = scores.masked_fill(~attention_mask, float("-inf")) else: scores = scores + attention_mask.float() attn_weights = F.softmax(scores, dim=-1).nan_to_num().to(q.dtype) # (B, n_heads, T, head_dim) -> (B, T, n_heads * head_dim) attn_output = attn_output.transpose(1, 2).contiguous() # Gated attention: output * sigmoid(g_proj(input)) if self.config.use_gated_attention: gate = self.g_proj(hidden_states) if self.config.gate_full_precision: gate = gate.float() gate = torch.sigmoid(gate).to(attn_output.dtype) attn_output = attn_output.view(bsz, q_len, -1) * gate else: attn_output = attn_output.view(bsz, q_len, -1) # Output projection attn_output = self.o_proj(attn_output) return attn_output, attn_weights, past_key_value def _mlp_chunk_size() -> int: """Row budget for the chunked MLP forward. `0` (or negative) disables it.""" try: return int(os.environ.get("RIPPLE_MLP_CHUNK", "2048")) except ValueError: return 2048 class RicursiveMLP(nn.Module): def __init__(self, config: RicursiveConfig): super().__init__() self.clamp_threshold = config.ffn_clamp_threshold self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False) self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) def _forward_rows(self, x: torch.Tensor) -> torch.Tensor: gate = self.gate_proj(x) up = self.up_proj(x) if self.clamp_threshold: gate = gate.clamp(min=None, max=self.clamp_threshold) up = up.clamp(min=-self.clamp_threshold, max=self.clamp_threshold) return self.down_proj(F.silu(gate) * up) def forward(self, x: torch.Tensor) -> torch.Tensor: # Ripple's MLP expands hidden_size -> intermediate_size (5120 -> 40960 # for the 65B checkpoint), so a long prefill materialises `gate`, `up` # and `silu(gate) * up` at once. Every op here is row-wise, so splitting # the row dimension is exact: each row still uses the same kernels with # the same reduction order, and the pieces are concatenated back. Only # the transient peak shrinks. `RIPPLE_MLP_CHUNK` overrides the # row budget; `0` disables chunking. chunk = _mlp_chunk_size() if chunk <= 0 or not torch.is_tensor(x) or x.dim() not in (2, 3): return self._forward_rows(x) shape = x.shape hidden = shape[-1] rows = 1 for size in shape[:-1]: rows *= size if rows <= chunk: return self._forward_rows(x) flat = x.reshape(rows, hidden) pieces = [ self._forward_rows(flat[start : start + chunk]) for start in range(0, rows, chunk) ] out = torch.cat(pieces, dim=0) if len(pieces) > 1 else pieces[0] return out.reshape(*shape[:-1], out.shape[-1]) class RicursiveDecoderLayer(nn.Module): """PeriNorm decoder layer with 4 layer norms.""" def __init__(self, config: RicursiveConfig, layer_idx: int): super().__init__() self.self_attn = RicursiveAttention(config, layer_idx) self.mlp = RicursiveMLP(config) # PeriNorm: 4 norms self.input_layernorm = RicursiveRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = RicursiveRMSNorm( config.hidden_size, eps=config.rms_norm_eps ) self.pre_feedforward_layernorm = RicursiveRMSNorm( config.hidden_size, eps=config.rms_norm_eps ) self.post_feedforward_layernorm = RicursiveRMSNorm( config.hidden_size, eps=config.rms_norm_eps ) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, value_residual: Optional[torch.Tensor] = None, **kwargs, ) -> Tuple[torch.Tensor, ...]: # Attention sub-block (PeriNorm: pre-norm -> attn -> post-norm -> residual) normed = self.input_layernorm(hidden_states) attn_output, attn_weights, present_key_value = self.self_attn( normed, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, value_residual=value_residual, **kwargs, ) hidden_states = hidden_states + self.post_attention_layernorm(attn_output) # FFN sub-block (PeriNorm: pre-norm -> ffn -> post-norm -> residual) normed = self.pre_feedforward_layernorm(hidden_states) ff_output = self.mlp(normed) hidden_states = hidden_states + self.post_feedforward_layernorm(ff_output) outputs = (hidden_states,) if output_attentions: outputs += (attn_weights,) if use_cache: outputs += (present_key_value,) return outputs class RicursivePreTrainedModel(PreTrainedModel): config_class = RicursiveConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["RicursiveDecoderLayer"] _supports_sdpa = True _supports_cache_class = True class RicursiveModel(RicursivePreTrainedModel): def __init__(self, config: RicursiveConfig): super().__init__(config) self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size) self.layers = nn.ModuleList( [RicursiveDecoderLayer(config, i) for i in range(config.num_hidden_layers)] ) self.norm = RicursiveRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False self.post_init() def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = ( output_attentions if output_attentions is not None else self.config.output_attentions ) output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict checkpointing = self.gradient_checkpointing and self.training if checkpointing: if past_key_values is not None: raise ValueError("Gradient checkpointing does not support a supplied KV cache") # Recomputed layers must not append to a cache for a second time. use_cache = False if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) # embed_scale: non-folded checkpoints carry the scale (e.g. sqrt(hidden_size)) in the # config and apply it here at runtime; folded checkpoints set embed_scale=None and bake # the scale into embed_tokens.weight, so this is a no-op. This mirrors the SGLang native # port (sglang_ricursive) so served and trained forwards agree. It must run BEFORE the # value-residual seed below, whose source is the (scaled) layer-0 embeddings. embed_scale = getattr(self.config, "embed_scale", None) if embed_scale is not None: inputs_embeds = inputs_embeds * embed_scale if use_cache and past_key_values is None: past_key_values = _new_default_cache(self.config) if cache_position is None: past_seen_tokens = ( past_key_values.get_seq_length() if past_key_values is not None else 0 ) cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device, ) if position_ids is None: position_ids = cache_position.unsqueeze(0) # Build causal mask causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values ) hidden_states = inputs_embeds value_residual = hidden_states if self.config.use_valres else None all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None for layer in self.layers: if output_hidden_states: all_hidden_states += (hidden_states,) if checkpointing: # Pass the value-residual seed as an input too, so gradients # through the embedding path survive reentrant checkpointing. layer_outputs = self._gradient_checkpointing_func( partial(layer.__call__, **kwargs), hidden_states, causal_mask, position_ids, None, output_attentions, False, cache_position, value_residual, ) else: layer_outputs = layer( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, value_residual=value_residual, **kwargs, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = past_key_values if use_cache else None if not return_dict: return tuple( v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None ) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, ) def _update_causal_mask( self, attention_mask: Optional[torch.Tensor], input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Optional[Cache], ) -> Optional[torch.Tensor]: # When there's no explicit attention_mask (no padding) and we're doing # standard prefill or cached decoding, return None and let SDPA use # its efficient is_causal path. This also avoids numerical issues # from float min mask values with large activations. if attention_mask is None: return None if attention_mask.dim() == 4: # A caller-provided 4D mask already defines visibility (or additive # bias). Preserve it; each local layer still applies its window. dtype = torch.bool if attention_mask.dtype == torch.bool else input_tensor.dtype return attention_mask.to(device=input_tensor.device, dtype=dtype) if torch.all(attention_mask == 1): return None # Only build an explicit mask when we have a padding mask dtype, device = input_tensor.dtype, input_tensor.device sequence_length = input_tensor.shape[1] if cache_position is None: query_positions = torch.arange(sequence_length, device=device) else: query_positions = cache_position.to(device=device).reshape(-1) # Size the key axis from `cache_position`, which is exactly what the # attention layer uses to size its key tensor. # `past_key_values.get_max_cache_shape()` was removed in transformers 5, # and its replacement reports the *capacity* (e.g. `sliding_window`), # which disagrees with the live key count whenever the cache is not full. target_length = int(query_positions[-1].item()) + 1 target_length = max(target_length, attention_mask.shape[-1]) # A shorter padding mask leaves the trailing keys unspecified; treat them as # visible -- the causal term below still applies. if attention_mask.shape[-1] < target_length: pad = attention_mask.new_ones( attention_mask.shape[0], target_length - attention_mask.shape[-1] ) attention_mask = torch.cat([pad, attention_mask], dim=-1) key_positions = torch.arange(target_length, device=device) allowed = key_positions[None, :] <= query_positions[:, None] allowed = allowed[None, None, :, :] & attention_mask[:, None, None, :].to(torch.bool) causal_mask = torch.zeros( (attention_mask.shape[0], 1, sequence_length, target_length), dtype=dtype, device=device, ) # `0.0` for visible keys (not `1.0`, which would bias every score) and the # dtype minimum elsewhere. Building the mask from `allowed` also keeps a # single-token decode step from coming out fully masked. return causal_mask.masked_fill(~allowed, torch.finfo(dtype).min) def _cache_seq_length(past_key_values: Optional[Cache]) -> int: """Number of cached keys, across transformers 4.x and 5.x cache objects.""" if past_key_values is None: return 0 getter = getattr(past_key_values, "get_seq_length", None) if callable(getter): return int(getter()) try: return int(past_key_values[0][0].shape[-2]) except Exception: # pragma: no cover - unknown cache implementations return 0 def _cache_layer_seq_length( past_key_values: Optional[Cache], layer_idx: int ) -> Optional[int]: """Actual number of keys written for `layer_idx`, or `None` if unknown. `Cache.get_seq_length(layer_idx)` counts *written* tokens. For a preallocated static layer that is smaller than the tensor `update()` returns until the cache fills; for an evicting sliding-window layer it is larger, because the layer only returns a trailing window. """ if past_key_values is None: return None getter = getattr(past_key_values, "get_seq_length", None) if not callable(getter): return None try: return int(getter(layer_idx)) except TypeError: # transformers 4.x cache without a layer argument pass except Exception: # pragma: no cover - unknown cache implementations return None try: return int(getter()) except Exception: # pragma: no cover - unknown cache implementations return None class RicursiveForCausalLM(RicursivePreTrainedModel, GenerationMixin): # Transformers 5 reads this mapping; Transformers 4 iterates its target keys. _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"} # transformers 5 no longer injects `GenerationMixin` into remote-code classes, # so inherit it explicitly above. Without it `generate` is missing entirely and # `PreTrainedModel.__init__` never builds `generation_config`. # # This attribute marks the native transformers-5 compatibility fixes so that # downstream loaders can skip their runtime monkeypatch. _ricursive_tf5_native_fix = RICURSIVE_TF5_COMPAT def __init__(self, config: RicursiveConfig): super().__init__(config) self.model = RicursiveModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.post_init() def get_input_embeddings(self): return self.model.embed_tokens def set_input_embeddings(self, value): self.model.embed_tokens = value def get_output_embeddings(self): return self.lm_head def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Optional[int] = None, num_items_in_batch: Optional[torch.Tensor] = None, **kwargs, ) -> Union[Tuple, CausalLMOutputWithPast]: return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, **kwargs, ) hidden_states = outputs[0] # Honour `logits_to_keep` by slicing the LM head *input*. Projecting the # full `[batch, seq, vocab]` tensor only to discard all but the last row # costs several GiB of transient memory at prefill batch sizes. # # When `labels` are supplied the full sequence must be kept: the causal-LM # loss shifts logits and labels by one position, so trimming both to the # same short window leaves the shift empty (and the loss NaN). Labeled # calls therefore project the full sequence to get a defined loss, then # trim only the *returned* logits so the output shape still honours # `logits_to_keep`. The generation path, which never passes labels, keeps # the memory win from slicing the LM head input. keep = ( int(logits_to_keep) if logits_to_keep is not None and logits_to_keep > 0 else 0 ) if keep and labels is None and hidden_states.shape[1] > keep: hidden_states = hidden_states[:, -keep:, :] logits = self.lm_head(hidden_states) loss = None if labels is not None: shift_logits = logits[..., :-1, :].float().contiguous() shift_labels = labels[..., 1:].to(logits.device).contiguous() loss = F.cross_entropy( shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1), reduction="sum" if num_items_in_batch is not None else "mean", ) if num_items_in_batch is not None: # Trainer supplies the token count across all accumulated microbatches. loss = loss / torch.as_tensor(num_items_in_batch, device=loss.device) if keep: # Loss used the full sequence; only the returned tensor is trimmed. logits = logits[..., -keep:, :] if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def _update_model_kwargs_for_generation( self, outputs, model_kwargs, is_encoder_decoder=False, num_new_tokens=1 ): """Advance explicit positions when the generation backend leaves them unchanged.""" position_ids = model_kwargs.get("position_ids") cache_position = model_kwargs.get("cache_position") model_kwargs = super()._update_model_kwargs_for_generation( outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens, ) # Transformers 4 leaves positions unchanged; Transformers 5 already extends them. if position_ids is not None and model_kwargs.get("position_ids") is position_ids: offsets = torch.arange( 1, num_new_tokens + 1, device=position_ids.device, dtype=position_ids.dtype ) model_kwargs["position_ids"] = torch.cat( [position_ids, position_ids[:, -1:] + offsets], dim=-1 ) # Transformers 5 no longer maintains cache_position. Keep an explicitly # supplied vector in sync, without advancing it twice on Transformers 4. if cache_position is not None and model_kwargs.get("cache_position") is cache_position: if model_kwargs.get("use_cache", True): model_kwargs["cache_position"] = cache_position[-1:] + num_new_tokens else: offsets = torch.arange( 1, num_new_tokens + 1, device=cache_position.device, dtype=cache_position.dtype ) model_kwargs["cache_position"] = torch.cat([cache_position, cache_position[-1:] + offsets]) return model_kwargs def prepare_inputs_for_generation( self, input_ids: torch.LongTensor, past_key_values: Optional[Cache] = None, attention_mask: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, cache_position: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, position_ids: Optional[torch.LongTensor] = None, next_sequence_length: Optional[int] = None, is_first_iteration: Optional[bool] = None, **kwargs, ): """Prepare a full prompt or a new-only block for cached generation. With a warm cache and no explicit ``cache_position``, inputs longer than the cached prefix are full prompts; shorter or equal-length inputs are new-only blocks. For a longer new-only block, pass ``torch.arange(past_length, past_length + input_ids.shape[1])`` as ``cache_position``. Attention masks may be padded to cache capacity and therefore cannot identify which input layout the caller supplied. """ # transformers 5.x does not move remote-code inputs onto the model device for # this model, so a plain CPU `input_ids` reaches the embedding lookup and # raises. Its own generation helper performs this move, so mirror it here. device = getattr(self, "device", None) if device is not None and getattr(device, "type", None) != "meta": if input_ids is not None and input_ids.device != device: input_ids = input_ids.to(device) if attention_mask is not None and attention_mask.device != device: attention_mask = attention_mask.to(device) if inputs_embeds is not None and inputs_embeds.device != device: inputs_embeds = inputs_embeds.to(device) # transformers derives `position_ids` from the *caller's* tensors before # this hook runs, so a CPU prompt yields CPU positions even after the # move above. RoPE then multiplies them against CUDA frequencies and fails. if position_ids is not None and position_ids.device != device: position_ids = position_ids.to(device) past_length = _cache_seq_length(past_key_values) use_embeds = inputs_embeds is not None and ( is_first_iteration or past_length == 0 or input_ids is None or input_ids.numel() == 0 ) input_tensor = inputs_embeds if use_embeds else input_ids input_length = input_tensor.shape[1] if cache_position is None: # Inputs no longer than the cache are treated as new-only blocks. # Longer inputs default to full prompts containing the cached prefix. # Mask width/content cannot distinguish a longer new-only block from # a full prompt with a capacity-sized mask (even an all-visible one). # Callers supplying longer new-only blocks must pass cache_position. full_prompt = past_length > 0 and input_length > past_length new_len = ( next_sequence_length if next_sequence_length is not None else input_length - past_length if full_prompt else input_length ) cache_position = torch.arange( past_length, past_length + new_len, device=input_tensor.device ) else: cache_position = cache_position.to(input_tensor.device) # Only slice when the cache is warm and `input_ids` still carries cached # tokens. Slicing unconditionally turned a `None` cache position into # `input_ids[:, None]` and injected a bogus axis. new_len = cache_position.numel() sliced = past_length > 0 and input_length > new_len if sliced: # During continuation, input_ids may contain only the new prompt # and generated tokens. Absolute cache offsets cannot index that # shorter buffer. Both layouts end with the unprocessed tokens. input_tensor = input_tensor[:, -new_len:] if new_len else input_tensor[:, :0] if use_embeds: model_inputs = {"inputs_embeds": input_tensor} else: model_inputs = {"input_ids": input_tensor} sequence_length = input_tensor.shape[1] if attention_mask is not None and attention_mask.dim() == 2 and new_len: required_length = int(cache_position[-1]) + 1 if attention_mask.shape[-1] < required_length: # A current-block mask omits an implicitly visible cached prefix. attention_mask = F.pad(attention_mask, (required_length - attention_mask.shape[-1], 0), value=1) # Derive per-row positions from the padding mask. Using `cache_position` for # every row is only correct when nothing is padded: with left padding the real # tokens of a short row sit at higher indices, so a shared position vector # shifts its RoPE phases and the row decodes to garbage. if position_ids is not None: # Callers may supply positions for the full prompt/capacity, or may # already have selected just the positions for the current block. if position_ids.shape[-1] != sequence_length: if new_len and position_ids.shape[-1] > int(cache_position[-1]): position_ids = position_ids[:, cache_position] else: position_ids = position_ids[:, -sequence_length:] if sequence_length else position_ids[:, :0] elif attention_mask is not None and attention_mask.dim() == 2: position_ids = attention_mask.long().cumsum(-1) - 1 position_ids = position_ids.masked_fill(attention_mask == 0, 0) # The mask may include unused cache capacity after the current block. # Select its actual cache slots, not the final entries of that mask. position_ids = position_ids[:, cache_position] else: position_ids = cache_position.unsqueeze(0) model_inputs.update( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": use_cache, "attention_mask": attention_mask, "cache_position": cache_position, } ) # Generation output flags and model-specific forward options must reach # forward. Labels are never generation inputs. model_inputs.update({key: value for key, value in kwargs.items() if key != "labels"}) if use_cache is not False: # Generation only reads the final position. Requesting one row keeps # the 100k-way LM head from projecting the whole prompt. model_inputs.setdefault("logits_to_keep", 1) return model_inputs
"""HuggingFace modeling implementation for the Ricursive transformer architecture.
Designed for inference with vLLM (via --trust-remote-code). Key differences from
standard Llama/Mistral:
- PeriNorm blocks: 4 layer norms per block (pre/post attention + pre/post FFN)
- Gated attention: elementwise sigmoid gate on attention output
- Mixed local/global layers: local = sliding window + RoPE, global = full attention + no RoPE
- Per-head Q/K RMSNorm
Transformers 5.x compatibility (2026-11-05)
------------------------------------------
`generate()` used to break on the very first prefill because
`prepare_inputs_for_generation` sliced `input_ids` with a `None`
`cache_position`, and the causal/sliding-window mask builders assumed the
transformers 4.x cache API. The fixes are applied inline below:
* `prepare_inputs_for_generation` rebuilds a missing `cache_position`, slices
only when the cache is warm, derives per-row `position_ids` from the padding
mask, and moves inputs to the model device.
* `_update_causal_mask` sizes the key axis from `cache_position` instead of the
removed `get_max_cache_shape()`, emits `0.0` for visible keys, and never
returns an all-masked single-token decode row.
* `_apply_sliding_window_mask` anchors key positions at the first key the layer
actually holds (the cache evicts from the left) and is skipped when the
window cannot bind.
* `RicursiveForCausalLM` inherits `GenerationMixin` explicitly; transformers 5
no longer injects it into remote-code classes.
DFlash2 inference fixes (2026-11-06)
-----------------------------------
* `RicursiveAttention.forward` builds an end-anchored causal mask for a
multi-token verify block running against a warm cache. SDPA's implicit mask
is top-left aligned, so with `q_len < kv_len` it hid valid context.
Preallocated caches (`StaticCache`) return their whole capacity from
`update()`, so the mask is now built from the layer's *actual* cache length
and query positions; the unused tail is trimmed before attention.
* `RicursiveForCausalLM.forward` honours `logits_to_keep` by slicing the LM
head *input*, so the full `[batch, seq, vocab]` projection is never
materialised; `prepare_inputs_for_generation` requests one row. Labeled
calls keep the full sequence so the shifted causal-LM loss stays defined,
then trim only the returned logits to the requested width.
* `RicursiveMLP.forward` chunks the row dimension
(`RIPPLE_MLP_CHUNK`, default 2048, `0` disables) to bound the
prefill activation peak. Every op is row-wise, so the result is exact.
Cache-alignment fixes from review (2026-11-06)
----------------------------------------------
* A mask longer than the key tensor is now aligned to the absolute positions
the layer actually holds, `[total_kv - kv_len, total_kv)`, instead of to the
tail of the mask. With a mask sized to the allocated cache *capacity*
(`StaticCache`), the old `[-kv_len:]` slice selected unwritten slots after
sliding-window eviction and silently returned wrong attention.
* `RicursiveAttention.forward` forwards `cache_position` to
`past_key_value.update()`. A transformers 4.x `StaticCache` otherwise falls
back to `arange(q_len)` and the first decode step overwrites the prompt
instead of appending after it.
Model contract fixes (2026-11-09)
-------------------------------
* Generation supports embeddings-only prompts, continuation buffers, explicit
position advancement, short masks, and requested diagnostic outputs.
* Attention preserves caller 4D masks, reports requested weights without changing
SDPA outputs, and trims unused static capacity before expanding KV heads.
* Default caches follow the hybrid layer configuration on Transformers 5;
Transformers 4 keeps a plain, crop-compatible full cache. Training
implements gradient checkpointing and computes causal-LM loss in FP32.
* Loss honors accumulation-wide token counts; TP reshapes use local head counts.
* Tied embeddings support Transformers 4/5, and window masks preserve exclusions.
"""
import os
from functools import partial
from typing import Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_outputs import (
BaseModelOutputWithPast,
CausalLMOutputWithPast,
)
try: # transformers >= 4.26, re-exported from here in 5.x
from transformers.generation.utils import GenerationMixin
except ImportError: # pragma: no cover - very old transformers
from transformers.generation import GenerationMixin
from .configuration_ricursive import RicursiveConfig
# Set when the transformers-5.x `generate()` fixes above are present natively.
# Downstream loaders that monkeypatch older checkpoints check this marker and
# stand down instead of re-installing the equivalent runtime patch.
RICURSIVE_TF5_COMPAT = "2026-11-09.1"
def _transformers_major_version() -> int:
"""Return the installed Transformers major version, or 0 when unknown."""