xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-15 07:24:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix GLM4 model (#16618)

Browse Source 
Signed-off-by: intervitens <intervitens@tutanota.com>
This commit is contained in:
intervitens 2025-04-17 13:35:07 +03:00 committed by GitHub
parent d8e557b5e5
commit 5b1aca2ae3
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/models/supported_models.md
View File

@ -338,7 +338,7 @@ See [this page](#generative-models) for more information on how to use generativ
* ✅︎ * ✅︎
- * `Glm4ForCausalLM` - * `Glm4ForCausalLM`
* GLM-4-0414 * GLM-4-0414
* `THUDM/GLM-4-32B-Chat-0414`, etc. * `THUDM/GLM-4-32B-0414`, etc.
* ✅︎ * ✅︎
* ✅︎ * ✅︎
- * `GPT2LMHeadModel` - * `GPT2LMHeadModel`

2
tests/models/registry.py
View File

@ -147,7 +147,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
min_transformers_version="4.50"), min_transformers_version="4.50"),
"GlmForCausalLM": _HfExamplesInfo("THUDM/glm-4-9b-chat-hf"), "GlmForCausalLM": _HfExamplesInfo("THUDM/glm-4-9b-chat-hf"),
"Glm4ForCausalLM": _HfExamplesInfo( "Glm4ForCausalLM": _HfExamplesInfo(
"THUDM/GLM-4-32B-Chat-0414", "THUDM/GLM-4-32B-0414",
is_available_online=False, is_available_online=False,
min_transformers_version="4.52.dev0" min_transformers_version="4.52.dev0"
), ),

8
vllm/model_executor/models/glm4.py
View File

@ -82,7 +82,7 @@ class Glm4Attention(nn.Module):
partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5) partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
self.head_dim = head_dim or hidden_size // self.total_num_heads self.head_dim = head_dim or hidden_size // self.total_num_heads
self.rotary_dim = int(partial_rotary_factor * self.head_dim) self.rotary_dim = self.head_dim
self.q_size = self.num_heads * self.head_dim self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5 self.scaling = self.head_dim**-0.5
@ -110,6 +110,7 @@ class Glm4Attention(nn.Module):
base=self.rope_theta, base=self.rope_theta,
rope_scaling=rope_scaling, rope_scaling=rope_scaling,
partial_rotary_factor=partial_rotary_factor, partial_rotary_factor=partial_rotary_factor,
is_neox_style=False,
) )
self.attn = Attention(self.num_heads, self.attn = Attention(self.num_heads,
self.head_dim, self.head_dim,
@ -197,13 +198,12 @@ class Glm4DecoderLayer(nn.Module):
) )
hidden_states = self.post_self_attn_layernorm(hidden_states) hidden_states = self.post_self_attn_layernorm(hidden_states)
hidden_states = residual + hidden_states
# Fully Connected # Fully Connected
hidden_states = self.post_attention_layernorm(hidden_states, residual) hidden_states, residual = self.post_attention_layernorm(
hidden_states, residual)
hidden_states = self.mlp(hidden_states) hidden_states = self.mlp(hidden_states)
hidden_states = self.post_mlp_layernorm(hidden_states) hidden_states = self.post_mlp_layernorm(hidden_states)
hidden_states = residual + hidden_states
return hidden_states, residual return hidden_states, residual