biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Add Mistral Large 3 and Ministral 3 (#29757)

Browse Source 
Signed-off-by: Julien Denize <julien.denize@mistral.ai>
Signed-off-by: Julien Denize <40604584+juliendenize@users.noreply.github.com>
Signed-off-by: Mickael Seznec <mickael@mistral.ai>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Mickael Seznec <mickael@mistral.ai>
This commit is contained in:
Julien Denize
2025-12-02 11:29:00 +01:00
committed by GitHub
parent 8bbcf8b6e7
commit d8c6210eea
16 changed files with 724 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

146
vllm/model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_H200,dtype=fp8_w8a8.json Normal file
View File

@@ -0,0 +1,146 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 5
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 3
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 3
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
}
}

4
vllm/model_executor/layers/mla.py
View File

@@ -111,6 +111,7 @@ class MultiHeadLatentAttentionWrapper(CustomOp):
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
llama_4_scaling: torch.Tensor | None = None,
) -> torch.Tensor:
q_c = None
kv_lora = None
@@ -159,6 +160,9 @@ class MultiHeadLatentAttentionWrapper(CustomOp):
hidden_states, q_c, positions, self.indexer_rope_emb
)
if llama_4_scaling is not None:
q *= llama_4_scaling
attn_out = self.mla_attn(
q,
kv_c_normed,

2
vllm/model_executor/layers/rotary_embedding/__init__.py
View File

@@ -238,7 +238,7 @@ def get_rope(
dtype,
**extra_kwargs,
)
elif scaling_type == "deepseek_yarn":
elif scaling_type in ["deepseek_yarn", "deepseek_llama_scaling"]:
scaling_factor = rope_parameters["factor"]
original_max_position = rope_parameters["original_max_position_embeddings"]
# assert max_position == original_max_position * scaling_factor