| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | import torch.nn.functional as F |
| 6 | from torch import nn |
| 7 | from transformers import Glm5NextTextConfig |
| 8 | |
| 9 | from vllm.config import ( |
| 10 | CacheConfig, |
| 11 | VllmConfig, |
| 12 | ) |
| 13 | from vllm.distributed import ( |
| 14 | get_tensor_model_parallel_world_size, |
| 15 | ) |
| 16 | from vllm.logger import init_logger |
| 17 | from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm |
| 18 | from vllm.model_executor.layers.linear import ( |
| 19 | ColumnParallelLinear, |
| 20 | MergedColumnParallelLinear, |
| 21 | ReplicatedLinear, |
| 22 | RowParallelLinear, |
| 23 | ) |
| 24 | from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper |
| 25 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 26 | from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope |
| 27 | from vllm.model_executor.models.deepseek_v2 import ( |
| 28 | DeepSeekV2FusedQkvAProjLinear, |
| 29 | DeepseekV32IndexerCache, |
| 30 | yarn_get_mscale, |
| 31 | ) |
| 32 | from vllm.model_executor.utils import maybe_disable_graph_partition |
| 33 | from vllm.models.glm5next.nvidia.ops.kpool_compress import fwht128_quant_fp8 |
| 34 | from vllm.models.glm5next.sparse_indexer import SparseAttnIndexerKpool |
| 35 | from vllm.platforms import current_platform |
| 36 | from vllm.utils.deep_gemm import PAGED_MQA_PAGE_SIZES |
| 37 | from vllm.utils.math_utils import cdiv, next_power_of_2 |
| 38 | from vllm.v1.kv_cache_interface import KpoolTailSpec, MLAAttentionSpec |
| 39 | |
| 40 | logger = init_logger(__name__) |
| 41 | |
| 42 | # Shared torch.compile config for the indexer's small-kernel leaves. The MLA |
| 43 | # indexer runs under breakable-CG (CompilationMode.NONE), which blocks FX-graph |
| 44 | # fusion of the surrounding eager ops; carving each cluster into its own |
| 45 | # @torch.compile leaf (backend==inductor) still fuses them. Matches the |
| 46 | # grouped_topk / _cast_sigmoid leaf pattern. |
| 47 | _INDEXER_COMPILE = dict( |
| 48 | dynamic=True, |
| 49 | backend=current_platform.simple_compile_backend, |
| 50 | options=maybe_disable_graph_partition(current_platform.simple_compile_backend), |
| 51 | ) |
| 52 | |
| 53 | |
| 54 | @torch.compile(**_INDEXER_COMPILE) |
| 55 | def _fused_indexer_k_norm( |
| 56 | x: torch.Tensor, weight: torch.Tensor, bias: torch.Tensor, dim: int, eps: float |
| 57 | ) -> torch.Tensor: |
| 58 | # Fuse fp32 cast + layer_norm + cast-back (was 3 kernels) into one. |
| 59 | return F.layer_norm(x.float(), (dim,), weight, bias, eps).type_as(x) |
| 60 | |
| 61 | |
| 62 | @torch.compile(**_INDEXER_COMPILE) |
| 63 | def _fused_indexer_weight_scale( |
| 64 | weights: torch.Tensor, q_scale: torch.Tensor, scale: float |
| 65 | ) -> torch.Tensor: |
| 66 | # Fuse the weight-scaling muls (was 2 kernels) into one. `scale` folds |
| 67 | # softmax_scale (head_dim**-0.5) and n_head**-0.5 into a single constant. |
| 68 | return (weights.unsqueeze(-1) * q_scale * scale).squeeze(-1) |
| 69 | |
| 70 | |
| 71 | @torch.compile(**_INDEXER_COMPILE) |
| 72 | def _pad_indexer_heads(x: torch.Tensor, pad: int) -> torch.Tensor: |
| 73 | # DeepGEMM MQA-logits needs num_heads in {32,64}; zero-pad the head dim. |
| 74 | # Fuse new_zeros + cat (was 2 kernels) into one. Pad values are zero (exact |
| 75 | # in fp8 e4m3 and zero-weight in the logits sum), so numerically a no-op. |
| 76 | return torch.cat([x, x.new_zeros(x.shape[0], pad, *x.shape[2:])], dim=1) |
| 77 | |
| 78 | |
| 79 | class Glm5NextIndexerCache(DeepseekV32IndexerCache): |
| 80 | """Indexer K cache that stores kpool-compressed entries. |
| 81 | |
| 82 | Setting ``tokens_per_state = index_kpool`` on the KV cache spec makes vLLM's |
| 83 | indexer metadata builder emit pool-granular ``slot_mapping`` / |
| 84 | ``seq_lens`` / ``cu_seq_lens`` / ``page_table`` for free, and shrinks the |
| 85 | cache allocation store one state per ``index_kpool`` tokens. The pool |
| 86 | *content* (softmax-weighted sum vs keep-every-Nth) is computed by the |
| 87 | kpool compress kernel inside the indexer op — the cache only provides the |
| 88 | addressing, which is identical for both schemes. |
| 89 | |
| 90 | The indexer shares one block with the co-located MLA (a single |
| 91 | ``MLAAttentionSpec`` / block_table), so ``block_size`` is the model-wide |
| 92 | ``cache_config.block_size``. DeepGEMM's paged-MQA kernel |
| 93 | (``csrc/apis/attention.hpp``) requires ``block_kv`` to be exactly 32 or |
| 94 | 64, so the storage block is virtually split into pool pages of the |
| 95 | largest such size that tiles it (``storage_kernel_block_size``); this |
| 96 | needs ``block_size`` to be a multiple of ``index_kpool * 32`` (512 for |
| 97 | ``index_kpool = 16``). A smaller block (e.g. the default 64) silently |
| 98 | collapses ``storage_block_size`` (64 // 16 = 4) and only fails later at |
| 99 | the opaque C++ assert; ``get_kv_cache_spec`` guards this up front |
| 100 | instead. |
| 101 | """ |
| 102 | |
| 103 | def __init__( |
| 104 | self, |
| 105 | *, |
| 106 | head_dim: int, |
| 107 | dtype: torch.dtype, |
| 108 | prefix: str, |
| 109 | cache_config, |
| 110 | index_kpool: int, |
| 111 | ): |
| 112 | super().__init__( |
| 113 | head_dim=head_dim, dtype=dtype, prefix=prefix, cache_config=cache_config |
| 114 | ) |
| 115 | assert index_kpool > 1, "Glm5NextIndexerCache expects index_kpool > 1" |
| 116 | # Keep chunked-prefill boundaries aligned to complete pools. |
| 117 | assert cache_config.block_size % index_kpool == 0, ( |
| 118 | "Glm5NextIndexerCache: cache_config.block_size " |
| 119 | f"({cache_config.block_size}) must be a multiple of index_kpool " |
| 120 | f"({index_kpool}) so chunked-prefill boundaries stay pool-aligned." |
| 121 | ) |
| 122 | self._index_kpool = index_kpool |
| 123 | |
| 124 | def get_kv_cache_spec(self, vllm_config: VllmConfig): |
| 125 | from dataclasses import replace |
| 126 | |
| 127 | spec = super().get_kv_cache_spec(vllm_config) |
| 128 | # ``tokens_per_state`` is the KV-spec representation of kpool |
| 129 | # compression in the current cache-layout API. |
| 130 | assert isinstance(spec, MLAAttentionSpec) |
| 131 | spec = replace(spec, tokens_per_state=self._index_kpool) |
| 132 | |
| 133 | # DeepGEMM paged-MQA takes block_kv in {32, 64}; the storage block |
| 134 | # (= block_size // index_kpool) is virtually split into pool pages of |
| 135 | # the largest such size that tiles it, so it must be a multiple of 32. |
| 136 | storage_block_size = spec.block_size // self._index_kpool |
| 137 | assert ( |
| 138 | spec.block_size % self._index_kpool == 0 and storage_block_size % 32 == 0 |
| 139 | ), ( |
| 140 | "Glm5NextIndexerCache: kpool indexer requires cache block_size to " |
| 141 | f"be a multiple of index_kpool * 32 ({self._index_kpool * 32}) so " |
| 142 | "that DeepGEMM paged-MQA pool pages (32 or 64 entries) tile the " |
| 143 | f"storage block, got block_size={spec.block_size} -> " |
| 144 | f"storage_block_size={storage_block_size}." |
| 145 | ) |
| 146 | max_page_size = max(PAGED_MQA_PAGE_SIZES) |
| 147 | min_page_size = min(PAGED_MQA_PAGE_SIZES) |
| 148 | if storage_block_size <= max_page_size: |
| 149 | page_size = storage_block_size |
| 150 | elif storage_block_size % max_page_size == 0: |
| 151 | page_size = max_page_size |
| 152 | else: |
| 153 | page_size = min_page_size |
| 154 | return replace( |
| 155 | spec, |
| 156 | storage_block_size=page_size * self._index_kpool, |
| 157 | ) |
| 158 | |
| 159 | |
| 160 | class Glm5NextTailCache(DeepseekV32IndexerCache): |
| 161 | """Paged circular buffer for the kpool indexer's in-progress (tail) pool. |
| 162 | |
| 163 | Holds the trailing incomplete pool's raw K + gate score: one block of |
| 164 | ``ring`` slots per request, overwritten in place by ``pos % ring`` |
| 165 | as decode/spec-decode advances. Prefill seeds it (instead of discarding the |
| 166 | tail raw K+gate); the connector transfers it across PD; decode reads it to |
| 167 | compress the boundary pool correctly. ``KpoolTailSpec`` / |
| 168 | ``KpoolTailManager`` provide the no-prune, 1-block/req allocation that lets |
| 169 | the in-progress pool survive across steps and across transfer. |
| 170 | |
| 171 | Stores raw bf16 K (``head_dim``) as the "K" half of each block and the |
| 172 | bf16 gate score (``head_dim``) as the "V" half -- not the fp8-compressed |
| 173 | entry, which lives in ``Glm5NextIndexerCache``. |
| 174 | """ |
| 175 | |
| 176 | def __init__( |
| 177 | self, |
| 178 | *, |
| 179 | head_dim: int, |
| 180 | dtype: torch.dtype, |
| 181 | prefix: str, |
| 182 | cache_config, |
| 183 | index_kpool: int, |
| 184 | ): |
| 185 | super().__init__( |
| 186 | head_dim=head_dim, dtype=dtype, prefix=prefix, cache_config=cache_config |
| 187 | ) |
| 188 | assert index_kpool > 1, "Glm5NextTailCache expects index_kpool > 1" |
| 189 | self._index_kpool = index_kpool |
| 190 | |
| 191 | def get_kv_cache_spec(self, vllm_config: VllmConfig): |
| 192 | # The two head slots form [K, gate score] in the generic |
| 193 | # [block, head, state, content] cache view. |
| 194 | # Drafts are stashed before acceptance. With a one-pool ring, the |
| 195 | # drafts behind a rejected pool-completing draft overwrite the keys |
| 196 | # read by its redo. |
| 197 | span = self._index_kpool + vllm_config.num_speculative_tokens |
| 198 | ring = self._index_kpool * next_power_of_2(cdiv(span, self._index_kpool)) |
| 199 | # ring must divide the attention block size (a multiple of 128). |
| 200 | assert self.cache_config.block_size % ring == 0, ( |
| 201 | f"Glm5NextTailCache: cache_config.block_size " |
| 202 | f"({self.cache_config.block_size}) must be a multiple of the " |
| 203 | f"tail ring ({ring})" |
| 204 | ) |
| 205 | return KpoolTailSpec( |
| 206 | block_size=ring, |
| 207 | num_kv_heads=2, |
| 208 | head_size=self.head_dim, |
| 209 | head_size_v=0, |
| 210 | dtype=torch.bfloat16, |
| 211 | sliding_window=ring, |
| 212 | ) |
| 213 | |
| 214 | def get_attn_backend(self): |
| 215 | from vllm.v1.attention.backends.mla.indexer import KpoolTailBackend |
| 216 | |
| 217 | return KpoolTailBackend |
| 218 | |
| 219 | |
| 220 | class Indexer(nn.Module): |
| 221 | def __init__( |
| 222 | self, |
| 223 | vllm_config: VllmConfig, |
| 224 | config: Glm5NextTextConfig, |
| 225 | hidden_size: int, |
| 226 | q_lora_rank: int, |
| 227 | quant_config: QuantizationConfig | None, |
| 228 | cache_config: CacheConfig | None, |
| 229 | topk_indices_buffer: torch.Tensor | None, |
| 230 | prefix: str = "", |
| 231 | ): |
| 232 | super().__init__() |
| 233 | self.vllm_config = vllm_config |
| 234 | self.config = config |
| 235 | self.quant_config = quant_config |
| 236 | # self.indexer_cfg = config.attn_module_list_cfg[0]["attn_index"] |
| 237 | # Indexer is only constructed for v32 configs, where these sparse-indexer |
| 238 | # fields are guaranteed populated; narrow away the `int | None` declared |
| 239 | # on Glm5NextTextConfig for the optional-indexer case. |
| 240 | assert config.index_topk is not None |
| 241 | assert config.index_n_heads is not None |
| 242 | assert config.index_head_dim is not None |
| 243 | assert config.index_kpool is not None |
| 244 | self.topk_tokens = config.index_topk |
| 245 | self.n_head = config.index_n_heads # 64 |
| 246 | self.head_dim = config.index_head_dim # 128 |
| 247 | self.rope_dim = config.qk_rope_head_dim # 64 |
| 248 | self.index_kpool = config.index_kpool |
| 249 | self.q_lora_rank = q_lora_rank # 1536 |
| 250 | |
| 251 | # kpool |
| 252 | self.index_kpool_compress_ape = nn.Parameter( |
| 253 | torch.zeros(self.index_kpool, self.head_dim, dtype=torch.float32) |
| 254 | ) |
| 255 | # Keep the checkpoint name ``index_kpool_compress_gate`` without a |
| 256 | # ``.weight`` suffix. F.linear consumes its [head_dim, hidden_size] shape. |
| 257 | self.index_kpool_compress_gate = nn.Parameter( |
| 258 | torch.empty(self.head_dim, hidden_size, dtype=torch.bfloat16) |
| 259 | ) |
| 260 | |
| 261 | # no tensor parallel, just replicated |
| 262 | self.wq_b = ReplicatedLinear( |
| 263 | self.q_lora_rank, |
| 264 | self.head_dim * self.n_head, |
| 265 | bias=False, |
| 266 | quant_config=quant_config, |
| 267 | prefix=f"{prefix}.wq_b", |
| 268 | ) |
| 269 | # Fused wk + weights_proj: single GEMM producing [head_dim + n_head]. |
| 270 | # FP8 wk weights are upcasted to BF16 during loading to maintain fusion. |
| 271 | self.wk_weights_proj = MergedColumnParallelLinear( |
| 272 | hidden_size, |
| 273 | [self.head_dim, self.n_head], |
| 274 | bias=False, |
| 275 | quant_config=None, |
| 276 | disable_tp=True, |
| 277 | prefix=f"{prefix}.wk_weights_proj", |
| 278 | ) |
| 279 | self.k_norm = LayerNorm(self.head_dim, eps=1e-6, dtype=torch.float32) |
| 280 | self.softmax_scale = self.head_dim**-0.5 |
| 281 | |
| 282 | # Hadamard-128 rotation of the indexer query is fused with the FP8 |
| 283 | # quant (see forward: fwht128_quant_fp8) -- no precomputed matrix. |
| 284 | |
| 285 | self.scale_fmt = "ue8m0" |
| 286 | self.quant_block_size = 128 # TODO: get from config |
| 287 | self.topk_indices_buffer = topk_indices_buffer |
| 288 | self._wp_fp32: torch.Tensor | None = None |
| 289 | |
| 290 | # NOTE: (zyongye) we use fp8 naive cache, |
| 291 | # where we store value in fp8 and scale in fp32 |
| 292 | # per self.quant_block_size element |
| 293 | self.k_cache = Glm5NextIndexerCache( |
| 294 | head_dim=self.head_dim + self.head_dim // self.quant_block_size * 4, |
| 295 | dtype=torch.uint8, |
| 296 | prefix=f"{prefix}.k_cache", |
| 297 | cache_config=cache_config, |
| 298 | index_kpool=self.index_kpool, |
| 299 | ) |
| 300 | # Paged tail cache (in-progress pool's raw K + gate score). Written by |
| 301 | # prefill (seeds the boundary pool) and decode (per-step stash); read by |
| 302 | # the decode kernel to compress the boundary pool. Transferred across PD |
| 303 | # so the decode side sees the prefill tail. See KpoolTailSpec/Manager. |
| 304 | self.tail_cache = Glm5NextTailCache( |
| 305 | head_dim=self.head_dim, |
| 306 | dtype=torch.bfloat16, |
| 307 | prefix=f"{prefix}.tail_cache", |
| 308 | cache_config=cache_config, |
| 309 | index_kpool=self.index_kpool, |
| 310 | ) |
| 311 | self.max_pool_len = vllm_config.model_config.max_model_len // self.index_kpool |
| 312 | self.prefix = prefix |
| 313 | from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size |
| 314 | |
| 315 | self.max_total_seq_len = ( |
| 316 | get_max_prefill_buffer_size(vllm_config) // self.index_kpool |
| 317 | ) |
| 318 | self.indexer_op = SparseAttnIndexerKpool( |
| 319 | self.k_cache, |
| 320 | self.quant_block_size, |
| 321 | self.scale_fmt, |
| 322 | self.topk_tokens, |
| 323 | self.head_dim, |
| 324 | self.max_pool_len, |
| 325 | self.max_total_seq_len, |
| 326 | self.topk_indices_buffer, |
| 327 | tail_cache=self.tail_cache, |
| 328 | ) |
| 329 | |
| 330 | def forward( |
| 331 | self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb |
| 332 | ) -> torch.Tensor: |
| 333 | q, _ = self.wq_b(qr) |
| 334 | q = q.view(-1, self.n_head, self.head_dim) |
| 335 | |
| 336 | # Compute the head gate in fp32; bf16 error can change near-tie pool |
| 337 | # rankings on long-context tasks. Cache it after weights are loaded. |
| 338 | kw, _ = self.wk_weights_proj(hidden_states) |
| 339 | k = kw[:, : self.head_dim] |
| 340 | if self._wp_fp32 is None: |
| 341 | self._wp_fp32 = ( |
| 342 | self.wk_weights_proj.weight.data[self.head_dim :, :] |
| 343 | .t() |
| 344 | .contiguous() |
| 345 | .float() |
| 346 | ) |
| 347 | weights = torch.mm(hidden_states.float(), self._wp_fp32) |
| 348 | |
| 349 | k = _fused_indexer_k_norm( |
| 350 | k, self.k_norm.weight, self.k_norm.bias, self.head_dim, self.k_norm.eps |
| 351 | ) |
| 352 | |
| 353 | if self.rope_dim > 0: |
| 354 | q_pe, q_nope = torch.split( |
| 355 | q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 |
| 356 | ) |
| 357 | k_pe, k_nope = torch.split( |
| 358 | k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 |
| 359 | ) |
| 360 | |
| 361 | q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1)) |
| 362 | # Note: RoPE (NeoX) can introduce extra leading dimensions during |
| 363 | # compilation so we need to reshape back to token-flattened shapes |
| 364 | q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim) |
| 365 | k_pe = k_pe.reshape(-1, 1, self.rope_dim) |
| 366 | |
| 367 | # `rotary_emb` is shape-preserving; `q_pe` is already |
| 368 | # [num_tokens, n_head, rope_dim]. |
| 369 | q = torch.cat([q_pe, q_nope], dim=-1) |
| 370 | # `k_pe` is [num_tokens, 1, rope_dim] (MQA). |
| 371 | k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1) |
| 372 | # else: qk_rope_head_dim=0 — no rope component. q is already |
| 373 | # [num_tokens, n_head, head_dim] and k is [num_tokens, head_dim] (all |
| 374 | # nope), so skip the rope split / rotary / cat entirely; otherwise the |
| 375 | # split/reshape would build 0-element tensors (breaks dynamo tracing). |
| 376 | |
| 377 | # Rotate Q into the cached K basis before computing fp8 MQA logits. |
| 378 | # Fusing the fp32 FWHT and quantization avoids an intermediate HBM |
| 379 | # round-trip and bf16 matrix-rounding bias. |
| 380 | assert self.head_dim == 128 and self.quant_block_size == 128 |
| 381 | assert self.scale_fmt == "ue8m0" |
| 382 | q = q.view(-1, self.head_dim) |
| 383 | q_fp8, q_scale = fwht128_quant_fp8(q) |
| 384 | q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim) |
| 385 | q_scale = q_scale.view(-1, self.n_head, 1) |
| 386 | |
| 387 | weights = _fused_indexer_weight_scale( |
| 388 | weights, q_scale, self.softmax_scale * self.n_head**-0.5 |
| 389 | ) |
| 390 | |
| 391 | # kpool: per-token gate score driving the softmax-weighted pool. Computed |
| 392 | # from the same hidden_states that produced `k`, so it stays token-aligned. |
| 393 | # F.linear(x, gate) = x @ gate.T with gate [head_dim, hidden_size]. |
| 394 | gate_score = F.linear(hidden_states, self.index_kpool_compress_gate) |
| 395 | |
| 396 | # DeepGEMM's MQA-logits kernels (fp8_mqa_logits / |
| 397 | # fp8_fp4_paged_mqa_logits) require num_heads in {32, 64}; this |
| 398 | # checkpoint uses index_n_heads=16. Zero-pad q and the per-head |
| 399 | # weights: logits are a weights-weighted sum over heads, so |
| 400 | # zero-weight padded heads contribute exactly nothing. |
| 401 | if self.n_head < 32: |
| 402 | pad = 32 - self.n_head |
| 403 | q_fp8 = _pad_indexer_heads(q_fp8, pad) |
| 404 | weights = _pad_indexer_heads(weights, pad) |
| 405 | |
| 406 | return self.indexer_op( |
| 407 | hidden_states, |
| 408 | q_fp8, |
| 409 | k, |
| 410 | weights, |
| 411 | gate_score=gate_score, |
| 412 | compress_ape=self.index_kpool_compress_ape, |
| 413 | index_kpool=self.index_kpool, |
| 414 | positions=positions, |
| 415 | ) |
| 416 | |
| 417 | |
| 418 | class Glm5NextMLAAttention(nn.Module): |
| 419 | def __init__( |
| 420 | self, |
| 421 | vllm_config: VllmConfig, |
| 422 | config: Glm5NextTextConfig, |
| 423 | hidden_size: int, |
| 424 | num_heads: int, |
| 425 | qk_nope_head_dim: int, |
| 426 | qk_rope_head_dim: int, |
| 427 | v_head_dim: int, |
| 428 | q_lora_rank: int | None, |
| 429 | kv_lora_rank: int, |
| 430 | max_position_embeddings: int = 8192, |
| 431 | cache_config: CacheConfig | None = None, |
| 432 | quant_config: QuantizationConfig | None = None, |
| 433 | prefix: str = "", |
| 434 | topk_indices_buffer: torch.Tensor | None = None, |
| 435 | input_size: int | None = None, |
| 436 | skip_rope: bool | None = False, |
| 437 | ) -> None: |
| 438 | super().__init__() |
| 439 | self.hidden_size = hidden_size |
| 440 | self.qk_nope_head_dim = qk_nope_head_dim |
| 441 | self.qk_rope_head_dim = qk_rope_head_dim |
| 442 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 443 | self.v_head_dim = v_head_dim |
| 444 | |
| 445 | self.q_lora_rank = q_lora_rank |
| 446 | self.kv_lora_rank = kv_lora_rank |
| 447 | |
| 448 | self.num_heads = num_heads |
| 449 | tp_size = get_tensor_model_parallel_world_size() |
| 450 | assert num_heads % tp_size == 0 |
| 451 | self.num_local_heads = num_heads // tp_size |
| 452 | |
| 453 | self.scaling = self.qk_head_dim**-0.5 |
| 454 | self.max_position_embeddings = max_position_embeddings |
| 455 | |
| 456 | # Use input_size for projection input dimensions if provided, |
| 457 | # otherwise default to hidden_size (used in Eagle3 Deepseek with MLA) |
| 458 | proj_input_size = input_size if input_size is not None else self.hidden_size |
| 459 | |
| 460 | if self.q_lora_rank is not None: |
| 461 | self.fused_qkv_a_proj = DeepSeekV2FusedQkvAProjLinear( |
| 462 | proj_input_size, |
| 463 | [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 464 | quant_config=quant_config, |
| 465 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 466 | ) |
| 467 | else: |
| 468 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 469 | proj_input_size, |
| 470 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 471 | bias=False, |
| 472 | quant_config=quant_config, |
| 473 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 474 | ) |
| 475 | |
| 476 | if self.q_lora_rank is not None: |
| 477 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 478 | self.q_b_proj = ColumnParallelLinear( |
| 479 | self.q_lora_rank, |
| 480 | self.num_heads * self.qk_head_dim, |
| 481 | bias=False, |
| 482 | quant_config=quant_config, |
| 483 | prefix=f"{prefix}.q_b_proj", |
| 484 | ) |
| 485 | else: |
| 486 | self.q_proj = ColumnParallelLinear( |
| 487 | proj_input_size, |
| 488 | self.num_heads * self.qk_head_dim, |
| 489 | bias=False, |
| 490 | quant_config=quant_config, |
| 491 | prefix=f"{prefix}.q_proj", |
| 492 | ) |
| 493 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 494 | self.kv_b_proj = ColumnParallelLinear( |
| 495 | self.kv_lora_rank, |
| 496 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 497 | bias=False, |
| 498 | quant_config=quant_config, |
| 499 | prefix=f"{prefix}.kv_b_proj", |
| 500 | ) |
| 501 | self.o_proj = RowParallelLinear( |
| 502 | self.num_heads * self.v_head_dim, |
| 503 | self.hidden_size, |
| 504 | bias=False, |
| 505 | quant_config=quant_config, |
| 506 | prefix=f"{prefix}.o_proj", |
| 507 | ) |
| 508 | |
| 509 | # `Glm5NextTextConfig` is NoPE-only and so declares no `rope_parameters`; |
| 510 | # a checkpoint may still ship one. |
| 511 | rope_parameters = getattr(config, "rope_parameters", None) |
| 512 | |
| 513 | if not skip_rope: |
| 514 | assert rope_parameters is not None |
| 515 | if rope_parameters["rope_type"] != "default": |
| 516 | rope_parameters["rope_type"] = ( |
| 517 | "deepseek_llama_scaling" |
| 518 | if rope_parameters.get("attention_factor") == 1.0 |
| 519 | else "deepseek_yarn" |
| 520 | ) |
| 521 | |
| 522 | self.rotary_emb: RotaryEmbedding | None = get_rope( |
| 523 | qk_rope_head_dim, |
| 524 | max_position=max_position_embeddings, |
| 525 | rope_parameters=rope_parameters, |
| 526 | is_neox_style=False, |
| 527 | ) |
| 528 | |
| 529 | if ( |
| 530 | rope_parameters["rope_type"] != "default" |
| 531 | and rope_parameters["rope_type"] == "deepseek_yarn" |
| 532 | ): |
| 533 | mscale_all_dim = rope_parameters.get("mscale_all_dim", False) |
| 534 | scaling_factor = rope_parameters["factor"] |
| 535 | mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) |
| 536 | self.scaling = self.scaling * mscale * mscale |
| 537 | else: |
| 538 | self.rotary_emb = None |
| 539 | |
| 540 | self.is_v32 = config.index_topk is not None |
| 541 | |
| 542 | if self.is_v32: |
| 543 | self.indexer_rope_emb: RotaryEmbedding | None = get_rope( |
| 544 | qk_rope_head_dim, |
| 545 | max_position=max_position_embeddings, |
| 546 | rope_parameters=rope_parameters, |
| 547 | is_neox_style=not config.indexer_rope_interleave, |
| 548 | ) |
| 549 | # The sparse indexer projects from the MLA q-lora rank, which is |
| 550 | # always set for v32 MLA configs; narrow away the `int | None`. |
| 551 | assert q_lora_rank is not None |
| 552 | self.indexer: Indexer | None = Indexer( |
| 553 | vllm_config, |
| 554 | config, |
| 555 | hidden_size, |
| 556 | q_lora_rank, |
| 557 | quant_config, |
| 558 | cache_config, |
| 559 | topk_indices_buffer, |
| 560 | f"{prefix}.indexer", |
| 561 | ) |
| 562 | |
| 563 | else: |
| 564 | self.indexer_rope_emb = None |
| 565 | self.indexer = None |
| 566 | |
| 567 | mla_modules = MLAModules( |
| 568 | kv_a_layernorm=self.kv_a_layernorm, |
| 569 | kv_b_proj=self.kv_b_proj, |
| 570 | rotary_emb=self.rotary_emb, |
| 571 | o_proj=self.o_proj, |
| 572 | fused_qkv_a_proj=self.fused_qkv_a_proj |
| 573 | if self.q_lora_rank is not None |
| 574 | else None, |
| 575 | kv_a_proj_with_mqa=self.kv_a_proj_with_mqa |
| 576 | if self.q_lora_rank is None |
| 577 | else None, |
| 578 | q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None, |
| 579 | q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None, |
| 580 | q_proj=self.q_proj if self.q_lora_rank is None else None, |
| 581 | indexer=self.indexer, |
| 582 | indexer_rotary_emb=self.indexer_rope_emb, |
| 583 | is_sparse=self.is_v32, |
| 584 | topk_indices_buffer=topk_indices_buffer, |
| 585 | ) |
| 586 | |
| 587 | self.mla_attn = MultiHeadLatentAttentionWrapper( |
| 588 | self.hidden_size, |
| 589 | self.num_local_heads, |
| 590 | self.scaling, |
| 591 | self.qk_nope_head_dim, |
| 592 | self.qk_rope_head_dim, |
| 593 | self.v_head_dim, |
| 594 | self.q_lora_rank, |
| 595 | self.kv_lora_rank, |
| 596 | mla_modules, |
| 597 | cache_config, |
| 598 | quant_config, |
| 599 | prefix, |
| 600 | skip_topk=False, |
| 601 | fuse_qkv_rmsnorm=True, |
| 602 | ) |
| 603 | |
| 604 | def forward( |
| 605 | self, hidden_states: torch.Tensor, positions: torch.Tensor |
| 606 | ) -> torch.Tensor: |
| 607 | # The wrapper also runs the sparse indexer before MLA attention. |
| 608 | return self.mla_attn(positions, hidden_states) |