From 80d6d557f307448fdad46995b5e237817b31144e Mon Sep 17 00:00:00 2001 From: Harry Mellor <19981378+hmellor@users.noreply.github.com> Date: Thu, 13 Aug 2026 16:10:09 +0100 Subject: [PATCH] Standardise weight tying on `ParallelLMHead.tie_weights` (#52147) Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) --- vllm/model_executor/models/arctic.py | 2 +- vllm/model_executor/models/bailing_moe.py | 15 +++++++-------- vllm/model_executor/models/bloom.py | 13 ++++++------- vllm/model_executor/models/chameleon.py | 2 +- vllm/model_executor/models/ernie45_moe.py | 2 +- vllm/model_executor/models/ernie45_vl_moe.py | 2 +- vllm/model_executor/models/ernie_mtp.py | 2 +- vllm/model_executor/models/exaone.py | 2 +- vllm/model_executor/models/exaone4.py | 2 +- vllm/model_executor/models/exaone4_5_mtp.py | 2 +- vllm/model_executor/models/exaone_moe.py | 2 +- vllm/model_executor/models/exaone_moe_mtp.py | 2 +- vllm/model_executor/models/falcon.py | 15 +++++++-------- vllm/model_executor/models/funasr.py | 15 +++++++-------- vllm/model_executor/models/gemma4_mtp.py | 2 +- vllm/model_executor/models/glm4.py | 15 +++++++-------- vllm/model_executor/models/gpt_neox.py | 2 +- vllm/model_executor/models/granite.py | 2 +- vllm/model_executor/models/granite4_vision.py | 2 +- vllm/model_executor/models/granitemoe.py | 2 +- vllm/model_executor/models/granitemoehybrid.py | 2 +- vllm/model_executor/models/granitemoeshared.py | 2 +- vllm/model_executor/models/hrm_text.py | 15 +++++++-------- vllm/model_executor/models/hunyuan_v1.py | 2 +- vllm/model_executor/models/hy_v3.py | 2 +- vllm/model_executor/models/idefics3.py | 2 +- vllm/model_executor/models/internlm2.py | 2 +- vllm/model_executor/models/interns1_pro.py | 2 +- vllm/model_executor/models/interns2_mobius.py | 15 +++++++-------- vllm/model_executor/models/iquest_loopcoder.py | 15 +++++++-------- vllm/model_executor/models/mamba.py | 13 ++++++------- vllm/model_executor/models/mellum.py | 2 +- vllm/model_executor/models/mimo.py | 15 +++++++-------- vllm/model_executor/models/mixtral.py | 2 +- vllm/model_executor/models/molmo.py | 15 +++++++-------- vllm/model_executor/models/moss_audio.py | 15 +++++++-------- vllm/model_executor/models/nemotron.py | 2 +- vllm/model_executor/models/olmo3.py | 15 +++++++-------- vllm/model_executor/models/olmo_hybrid.py | 15 +++++++-------- vllm/model_executor/models/openpangu.py | 2 +- vllm/model_executor/models/opt.py | 13 ++++++------- vllm/model_executor/models/orion.py | 2 +- vllm/model_executor/models/param2moe.py | 15 +++++++-------- vllm/model_executor/models/qwen2.py | 15 +++++++-------- vllm/model_executor/models/qwen2_moe.py | 2 +- vllm/model_executor/models/qwen3.py | 15 +++++++-------- vllm/model_executor/models/qwen3_5.py | 15 +++++++-------- vllm/model_executor/models/qwen3_5_mtp.py | 15 +++++++-------- vllm/model_executor/models/qwen3_moe.py | 2 +- .../models/qwen3_omni_moe_thinker.py | 2 +- vllm/model_executor/models/qwen3_vl.py | 15 +++++++-------- vllm/model_executor/models/qwen3_vl_moe.py | 2 +- vllm/model_executor/models/sarvam.py | 15 +++++++-------- vllm/model_executor/models/seed_oss.py | 15 +++++++-------- vllm/model_executor/models/solar.py | 2 +- vllm/model_executor/models/stablelm.py | 2 +- 56 files changed, 191 insertions(+), 214 deletions(-) diff --git a/vllm/model_executor/models/arctic.py b/vllm/model_executor/models/arctic.py index b7e8e6796f..ccb3f1976c 100644 --- a/vllm/model_executor/models/arctic.py +++ b/vllm/model_executor/models/arctic.py @@ -557,7 +557,7 @@ class ArcticForCausalLM(nn.Module, SupportsPP, SupportsQuant): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.num_experts = config.num_local_experts self.num_experts_per_tok = config.num_experts_per_tok diff --git a/vllm/model_executor/models/bailing_moe.py b/vllm/model_executor/models/bailing_moe.py index 0834de5ee5..a3da7e39e7 100644 --- a/vllm/model_executor/models/bailing_moe.py +++ b/vllm/model_executor/models/bailing_moe.py @@ -509,15 +509,14 @@ class BailingMoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): self.tie_word_embeddings = getattr(config, "tie_word_embeddings", False) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.tie_word_embeddings: - self.lm_head = self.model.word_embeddings - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.word_embeddings) self.logits_processor = LogitsProcessor(config.vocab_size) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/bloom.py b/vllm/model_executor/models/bloom.py index cdcf82f385..1347b0e605 100644 --- a/vllm/model_executor/models/bloom.py +++ b/vllm/model_executor/models/bloom.py @@ -324,14 +324,13 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant): self.transformer = BloomModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer") ) + self.lm_head = ParallelLMHead( + self.config.vocab_size, + self.config.hidden_size, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.config.tie_word_embeddings: - self.lm_head = self.transformer.word_embeddings - else: - self.lm_head = ParallelLMHead( - self.config.vocab_size, - self.config.hidden_size, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.transformer.word_embeddings) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( diff --git a/vllm/model_executor/models/chameleon.py b/vllm/model_executor/models/chameleon.py index 94b6b4d345..03c9fe1179 100644 --- a/vllm/model_executor/models/chameleon.py +++ b/vllm/model_executor/models/chameleon.py @@ -977,7 +977,7 @@ class ChameleonForConditionalGeneration( prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor(config.vocab_size, scale=logit_scale) diff --git a/vllm/model_executor/models/ernie45_moe.py b/vllm/model_executor/models/ernie45_moe.py index 89f0b32c66..5cc351a6aa 100644 --- a/vllm/model_executor/models/ernie45_moe.py +++ b/vllm/model_executor/models/ernie45_moe.py @@ -543,7 +543,7 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe self.lm_head = PPMissingLayer() if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/ernie45_vl_moe.py b/vllm/model_executor/models/ernie45_vl_moe.py index 5d36ca6d37..e18a3fec50 100644 --- a/vllm/model_executor/models/ernie45_vl_moe.py +++ b/vllm/model_executor/models/ernie45_vl_moe.py @@ -611,7 +611,7 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP): self.lm_head = PPMissingLayer() if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/ernie_mtp.py b/vllm/model_executor/models/ernie_mtp.py index b57da3698a..88aadbdb73 100644 --- a/vllm/model_executor/models/ernie_mtp.py +++ b/vllm/model_executor/models/ernie_mtp.py @@ -177,7 +177,7 @@ class ErnieMTP(nn.Module): ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: return self.model.embed_input_ids(input_ids) diff --git a/vllm/model_executor/models/exaone.py b/vllm/model_executor/models/exaone.py index 79314a7b93..0742ae870b 100644 --- a/vllm/model_executor/models/exaone.py +++ b/vllm/model_executor/models/exaone.py @@ -410,7 +410,7 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.transformer.wte.weight + self.lm_head = self.lm_head.tie_weights(self.transformer.wte) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/exaone4.py b/vllm/model_executor/models/exaone4.py index 8e8252f459..571aa3a426 100644 --- a/vllm/model_executor/models/exaone4.py +++ b/vllm/model_executor/models/exaone4.py @@ -410,7 +410,7 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/exaone4_5_mtp.py b/vllm/model_executor/models/exaone4_5_mtp.py index 7711f72e42..80e1903f9d 100644 --- a/vllm/model_executor/models/exaone4_5_mtp.py +++ b/vllm/model_executor/models/exaone4_5_mtp.py @@ -159,7 +159,7 @@ class Exaone4_5_MTP(ExaoneMoeMTP, SupportsMultiModal): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( self.unpadded_vocab_size, config.vocab_size ) diff --git a/vllm/model_executor/models/exaone_moe.py b/vllm/model_executor/models/exaone_moe.py index e942bf62ed..ef01330f0c 100644 --- a/vllm/model_executor/models/exaone_moe.py +++ b/vllm/model_executor/models/exaone_moe.py @@ -522,7 +522,7 @@ class ExaoneMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP): quant_config=quant_config, ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/exaone_moe_mtp.py b/vllm/model_executor/models/exaone_moe_mtp.py index ddd570d604..ebe1f8cbae 100644 --- a/vllm/model_executor/models/exaone_moe_mtp.py +++ b/vllm/model_executor/models/exaone_moe_mtp.py @@ -159,7 +159,7 @@ class ExaoneMoeMTP(nn.Module): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( self.unpadded_vocab_size, config.vocab_size ) diff --git a/vllm/model_executor/models/falcon.py b/vllm/model_executor/models/falcon.py index efd24b5144..cdc388e529 100644 --- a/vllm/model_executor/models/falcon.py +++ b/vllm/model_executor/models/falcon.py @@ -499,15 +499,14 @@ class FalconForCausalLM(nn.Module, SupportsPP): if config.tie_word_embeddings is not None else True ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.tie_word_embeddings: - self.lm_head = self.transformer.word_embeddings - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.transformer.word_embeddings) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.transformer.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/funasr.py b/vllm/model_executor/models/funasr.py index e128b22e8e..56b1b68f4f 100644 --- a/vllm/model_executor/models/funasr.py +++ b/vllm/model_executor/models/funasr.py @@ -919,15 +919,14 @@ class FunASRForConditionalGeneration( ) logit_scale = getattr(config, "logit_scale", 1.0) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.decoder.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.decoder.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size, scale=logit_scale) def get_language_model(self) -> torch.nn.Module: diff --git a/vllm/model_executor/models/gemma4_mtp.py b/vllm/model_executor/models/gemma4_mtp.py index 0fe972691e..9fd2b1f164 100644 --- a/vllm/model_executor/models/gemma4_mtp.py +++ b/vllm/model_executor/models/gemma4_mtp.py @@ -489,7 +489,7 @@ class Gemma4MTP(nn.Module): prefix=maybe_prefix(prefix, "lm_head"), ) if getattr(config, "tie_word_embeddings", True): - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( text_config.vocab_size, diff --git a/vllm/model_executor/models/glm4.py b/vllm/model_executor/models/glm4.py index a1fb94fb26..4c30e30008 100644 --- a/vllm/model_executor/models/glm4.py +++ b/vllm/model_executor/models/glm4.py @@ -252,15 +252,14 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/gpt_neox.py b/vllm/model_executor/models/gpt_neox.py index 907ab77760..8af7eb6792 100644 --- a/vllm/model_executor/models/gpt_neox.py +++ b/vllm/model_executor/models/gpt_neox.py @@ -285,7 +285,7 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP): prefix=maybe_prefix(prefix, "embed_out"), ) if self.config.tie_word_embeddings: - self.embed_out.weight = self.gpt_neox.embed_in.weight + self.embed_out = self.embed_out.tie_weights(self.gpt_neox.embed_in) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.gpt_neox.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/granite.py b/vllm/model_executor/models/granite.py index c46fefbf88..a0bc8350b4 100644 --- a/vllm/model_executor/models/granite.py +++ b/vllm/model_executor/models/granite.py @@ -365,7 +365,7 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) if hasattr(config, "logits_scaling"): diff --git a/vllm/model_executor/models/granite4_vision.py b/vllm/model_executor/models/granite4_vision.py index c6e4df2992..5c56717aa6 100644 --- a/vllm/model_executor/models/granite4_vision.py +++ b/vllm/model_executor/models/granite4_vision.py @@ -316,7 +316,7 @@ class Granite4VisionLLMForCausalLM(GraniteForCausalLM): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) if hasattr(config, "logits_scaling"): logit_scale /= config.logits_scaling diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py index 328ea9ce63..e1219e3337 100644 --- a/vllm/model_executor/models/granitemoe.py +++ b/vllm/model_executor/models/granitemoe.py @@ -505,7 +505,7 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( config.vocab_size, diff --git a/vllm/model_executor/models/granitemoehybrid.py b/vllm/model_executor/models/granitemoehybrid.py index a50a95a302..aa21c23f1e 100644 --- a/vllm/model_executor/models/granitemoehybrid.py +++ b/vllm/model_executor/models/granitemoehybrid.py @@ -663,7 +663,7 @@ class GraniteMoeHybridForCausalLM( prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( config.vocab_size, config.vocab_size, diff --git a/vllm/model_executor/models/granitemoeshared.py b/vllm/model_executor/models/granitemoeshared.py index 7abc682c58..603afc60d6 100644 --- a/vllm/model_executor/models/granitemoeshared.py +++ b/vllm/model_executor/models/granitemoeshared.py @@ -281,7 +281,7 @@ class GraniteMoeSharedForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor( config.vocab_size, diff --git a/vllm/model_executor/models/hrm_text.py b/vllm/model_executor/models/hrm_text.py index a7546b0cc4..dd7ba93e7b 100644 --- a/vllm/model_executor/models/hrm_text.py +++ b/vllm/model_executor/models/hrm_text.py @@ -487,15 +487,14 @@ class HrmTextForCausalLM(nn.Module): vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: diff --git a/vllm/model_executor/models/hunyuan_v1.py b/vllm/model_executor/models/hunyuan_v1.py index ef7c748519..b90d463627 100644 --- a/vllm/model_executor/models/hunyuan_v1.py +++ b/vllm/model_executor/models/hunyuan_v1.py @@ -927,7 +927,7 @@ class HunyuanV1ModelBase( prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/hy_v3.py b/vllm/model_executor/models/hy_v3.py index a4f4a558ec..0248ac3d9c 100644 --- a/vllm/model_executor/models/hy_v3.py +++ b/vllm/model_executor/models/hy_v3.py @@ -673,7 +673,7 @@ class HYV3ForCausalLM(nn.Module, SupportsPP, SupportsLoRA): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py index 732f5c021a..7f26e01c0c 100644 --- a/vllm/model_executor/models/idefics3.py +++ b/vllm/model_executor/models/idefics3.py @@ -539,7 +539,7 @@ class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsLo prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.text_config.tie_word_embeddings: - self.lm_head.weight = self.model.text_model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.text_model.embed_tokens) self.logits_processor = LogitsProcessor(config.text_config.vocab_size) def _parse_and_validate_image_input(self, **kwargs: object) -> ImageInputs | None: diff --git a/vllm/model_executor/models/internlm2.py b/vllm/model_executor/models/internlm2.py index 81487f9cad..ffc750c7c5 100644 --- a/vllm/model_executor/models/internlm2.py +++ b/vllm/model_executor/models/internlm2.py @@ -352,7 +352,7 @@ class InternLM2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SupportsQuant): prefix=maybe_prefix(prefix, "output"), ) if self.config.tie_word_embeddings: - self.output.weight = self.model.tok_embeddings.weight + self.output = self.output.tie_weights(self.model.tok_embeddings) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/interns1_pro.py b/vllm/model_executor/models/interns1_pro.py index 378177a557..e85ad87cdc 100644 --- a/vllm/model_executor/models/interns1_pro.py +++ b/vllm/model_executor/models/interns1_pro.py @@ -481,7 +481,7 @@ class InternS1ProMoeLLMForCausalLM(Qwen3MoeForCausalLM): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(self.config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/interns2_mobius.py b/vllm/model_executor/models/interns2_mobius.py index e13de751e6..1bbc614334 100644 --- a/vllm/model_executor/models/interns2_mobius.py +++ b/vllm/model_executor/models/interns2_mobius.py @@ -443,15 +443,14 @@ class InternS2MobiusForCausalLM(Qwen3_5ForCausalLMBase): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=self.quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=self.quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/iquest_loopcoder.py b/vllm/model_executor/models/iquest_loopcoder.py index 3755cba5d1..fdb90b962d 100644 --- a/vllm/model_executor/models/iquest_loopcoder.py +++ b/vllm/model_executor/models/iquest_loopcoder.py @@ -541,15 +541,14 @@ class IQuestLoopCoderForCausalLM(nn.Module): vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) diff --git a/vllm/model_executor/models/mamba.py b/vllm/model_executor/models/mamba.py index 6a77a58abf..72893c574f 100644 --- a/vllm/model_executor/models/mamba.py +++ b/vllm/model_executor/models/mamba.py @@ -188,14 +188,13 @@ class MambaForCausalLM( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "backbone") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.backbone.embeddings - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.backbone.embeddings) self.logits_processor = LogitsProcessor(config.vocab_size) diff --git a/vllm/model_executor/models/mellum.py b/vllm/model_executor/models/mellum.py index c20fa00e3d..2eb35687d7 100644 --- a/vllm/model_executor/models/mellum.py +++ b/vllm/model_executor/models/mellum.py @@ -221,7 +221,7 @@ class MellumForCausalLM(Qwen3MoeForCausalLM): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/mimo.py b/vllm/model_executor/models/mimo.py index e4247fa8d8..56250ba2f4 100644 --- a/vllm/model_executor/models/mimo.py +++ b/vllm/model_executor/models/mimo.py @@ -101,15 +101,14 @@ class MiMoForCausalLM(Qwen2ForCausalLM, nn.Module): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py index f9b452f47f..d154208606 100644 --- a/vllm/model_executor/models/mixtral.py +++ b/vllm/model_executor/models/mixtral.py @@ -403,7 +403,7 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py index cf4550a107..bdc46d83b9 100644 --- a/vllm/model_executor/models/molmo.py +++ b/vllm/model_executor/models/molmo.py @@ -1340,15 +1340,14 @@ class MolmoForCausalLM( self.img_patch_id = None + self.lm_head = ParallelLMHead( + config.embedding_size or config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.config.weight_tying: - self.lm_head = self.model.transformer.wte - else: - self.lm_head = ParallelLMHead( - config.embedding_size or config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.transformer.wte) self.logits_processor = LogitsProcessor( config.embedding_size or config.vocab_size diff --git a/vllm/model_executor/models/moss_audio.py b/vllm/model_executor/models/moss_audio.py index bef8057d2f..878eabc0e3 100644 --- a/vllm/model_executor/models/moss_audio.py +++ b/vllm/model_executor/models/moss_audio.py @@ -835,15 +835,14 @@ class MossQwen3ForCausalLM(Qwen3ForCausalLM): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: from .utils import PPMissingLayer diff --git a/vllm/model_executor/models/nemotron.py b/vllm/model_executor/models/nemotron.py index 6f0b61205b..97199c1841 100644 --- a/vllm/model_executor/models/nemotron.py +++ b/vllm/model_executor/models/nemotron.py @@ -404,7 +404,7 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/olmo3.py b/vllm/model_executor/models/olmo3.py index 922834a8ee..9ac4e61cf4 100644 --- a/vllm/model_executor/models/olmo3.py +++ b/vllm/model_executor/models/olmo3.py @@ -369,15 +369,14 @@ class Olmo3ForCausalLM(nn.Module, SupportsPP, SupportsLoRA): self.model = Olmo3Model( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=vllm_config.quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=vllm_config.quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/olmo_hybrid.py b/vllm/model_executor/models/olmo_hybrid.py index 51bc410363..122a1c1c5b 100644 --- a/vllm/model_executor/models/olmo_hybrid.py +++ b/vllm/model_executor/models/olmo_hybrid.py @@ -397,15 +397,14 @@ class OlmoHybridForCausalLM( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=vllm_config.quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=vllm_config.quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( diff --git a/vllm/model_executor/models/openpangu.py b/vllm/model_executor/models/openpangu.py index 3557bdd0a5..4a2030cfdd 100644 --- a/vllm/model_executor/models/openpangu.py +++ b/vllm/model_executor/models/openpangu.py @@ -1122,7 +1122,7 @@ class OpenPanguModelBase(nn.Module, SupportsPP, SupportsLoRA): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() self.logits_processor = LogitsProcessor(config.vocab_size) diff --git a/vllm/model_executor/models/opt.py b/vllm/model_executor/models/opt.py index 32bb532f5c..d78e51e3dd 100644 --- a/vllm/model_executor/models/opt.py +++ b/vllm/model_executor/models/opt.py @@ -349,14 +349,13 @@ class OPTForCausalLM(nn.Module, SupportsPP, SupportsLoRA): self.model = OPTModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) + self.lm_head = ParallelLMHead( + config.vocab_size, + config.word_embed_proj_dim, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.config.tie_word_embeddings: - self.lm_head = self.model.decoder.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.word_embed_proj_dim, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.decoder.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/orion.py b/vllm/model_executor/models/orion.py index 0871c347ac..2b58ee1e01 100644 --- a/vllm/model_executor/models/orion.py +++ b/vllm/model_executor/models/orion.py @@ -305,7 +305,7 @@ class OrionForCausalLM(nn.Module, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/param2moe.py b/vllm/model_executor/models/param2moe.py index 7830272c32..d865301a21 100644 --- a/vllm/model_executor/models/param2moe.py +++ b/vllm/model_executor/models/param2moe.py @@ -649,15 +649,14 @@ class Param2MoEForCausalLM( self.tie_word_embeddings: bool = getattr(config, "tie_word_embeddings", False) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen2.py b/vllm/model_executor/models/qwen2.py index f1d5f23a26..3ec8ac9319 100644 --- a/vllm/model_executor/models/qwen2.py +++ b/vllm/model_executor/models/qwen2.py @@ -460,15 +460,14 @@ class Qwen2ForCausalLM( ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py index 43274e4ba7..df1794d289 100644 --- a/vllm/model_executor/models/qwen2_moe.py +++ b/vllm/model_executor/models/qwen2_moe.py @@ -461,7 +461,7 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/qwen3.py b/vllm/model_executor/models/qwen3.py index cf136db5be..7db73a749f 100644 --- a/vllm/model_executor/models/qwen3.py +++ b/vllm/model_executor/models/qwen3.py @@ -295,15 +295,14 @@ class Qwen3ForCausalLM( ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen3_5.py b/vllm/model_executor/models/qwen3_5.py index 6ebf26c871..28c6a11899 100644 --- a/vllm/model_executor/models/qwen3_5.py +++ b/vllm/model_executor/models/qwen3_5.py @@ -334,15 +334,14 @@ class Qwen3_5ForCausalLMBase( ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=self.quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=self.quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen3_5_mtp.py b/vllm/model_executor/models/qwen3_5_mtp.py index 6421c8b60d..7dc7c53826 100644 --- a/vllm/model_executor/models/qwen3_5_mtp.py +++ b/vllm/model_executor/models/qwen3_5_mtp.py @@ -238,15 +238,14 @@ class Qwen3_5MTP(LocalArgmaxMixin, nn.Module, SupportsMultiModal): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=self.quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=self.quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen3_moe.py b/vllm/model_executor/models/qwen3_moe.py index c33166469a..5df982d741 100644 --- a/vllm/model_executor/models/qwen3_moe.py +++ b/vllm/model_executor/models/qwen3_moe.py @@ -575,7 +575,7 @@ class Qwen3MoeForCausalLM( prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/qwen3_omni_moe_thinker.py b/vllm/model_executor/models/qwen3_omni_moe_thinker.py index bb938167ae..2f84f12834 100755 --- a/vllm/model_executor/models/qwen3_omni_moe_thinker.py +++ b/vllm/model_executor/models/qwen3_omni_moe_thinker.py @@ -1111,7 +1111,7 @@ class Qwen3MoeLLMForCausalLM(Qwen3MoeForCausalLM): config.vocab_size, config.hidden_size, quant_config=quant_config ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py index c34d95db32..89495ff949 100644 --- a/vllm/model_executor/models/qwen3_vl.py +++ b/vllm/model_executor/models/qwen3_vl.py @@ -1707,15 +1707,14 @@ class Qwen3LLMForCausalLM(Qwen3ForCausalLM): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix="lm_head", + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix="lm_head", - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/qwen3_vl_moe.py b/vllm/model_executor/models/qwen3_vl_moe.py index 1360edfdd7..c7cc440b8c 100644 --- a/vllm/model_executor/models/qwen3_vl_moe.py +++ b/vllm/model_executor/models/qwen3_vl_moe.py @@ -168,7 +168,7 @@ class Qwen3MoeLLMForCausalLM(Qwen3MoeForCausalLM): prefix=maybe_prefix(prefix, "lm_head"), ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(self.config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors diff --git a/vllm/model_executor/models/sarvam.py b/vllm/model_executor/models/sarvam.py index 2b2fbfe708..04590a2a91 100644 --- a/vllm/model_executor/models/sarvam.py +++ b/vllm/model_executor/models/sarvam.py @@ -615,15 +615,14 @@ class SarvamMLAForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SarvamMixtureOfE self.tie_word_embeddings = getattr(config, "tie_word_embeddings", False) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if self.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/seed_oss.py b/vllm/model_executor/models/seed_oss.py index d2c767846d..f3fa8b2d5d 100644 --- a/vllm/model_executor/models/seed_oss.py +++ b/vllm/model_executor/models/seed_oss.py @@ -388,15 +388,14 @@ class SeedOssForCausalLM(nn.Module, SupportsLoRA, SupportsPP): ) if get_pp_group().is_last_rank: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=maybe_prefix(prefix, "lm_head"), + ) if config.tie_word_embeddings: - self.lm_head = self.model.embed_tokens - else: - self.lm_head = ParallelLMHead( - config.vocab_size, - config.hidden_size, - quant_config=quant_config, - prefix=maybe_prefix(prefix, "lm_head"), - ) + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) else: self.lm_head = PPMissingLayer() diff --git a/vllm/model_executor/models/solar.py b/vllm/model_executor/models/solar.py index 478a61da67..430337c77d 100644 --- a/vllm/model_executor/models/solar.py +++ b/vllm/model_executor/models/solar.py @@ -386,7 +386,7 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP): prefix=maybe_prefix(prefix, "lm_head"), ) if config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) logit_scale = getattr(config, "logit_scale", 1.0) self.logits_processor = LogitsProcessor( diff --git a/vllm/model_executor/models/stablelm.py b/vllm/model_executor/models/stablelm.py index 58758b11cd..b18f4051ca 100644 --- a/vllm/model_executor/models/stablelm.py +++ b/vllm/model_executor/models/stablelm.py @@ -294,7 +294,7 @@ class StablelmForCausalLM(nn.Module, SupportsPP): prefix=f"{prefix}.lm_head", ) if self.config.tie_word_embeddings: - self.lm_head.weight = self.model.embed_tokens.weight + self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors