[Model] Add LongCat-Flash (#23991)

Signed-off-by: yangxurui <yangxurui@meituan.com>
Co-authored-by: yangxurui <yangxurui@meituan.com>
This commit is contained in:
XuruiYang
2025-09-24 21:53:40 -07:00
committed by GitHub
co-authored by yangxurui
parent 90b139cfff
commit 845adb3ec6
31 changed files with 1357 additions and 66 deletions
+2 -2
View File
@@ -138,7 +138,7 @@ def test_flashinfer_per_tensor_moe_fp8_no_graph(
td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=True)
score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
topk_weights, topk_ids = FusedMoE.select_experts(
topk_weights, topk_ids, _ = FusedMoE.select_experts(
hidden_states=td.hidden_states,
router_logits=score,
use_grouped_topk=False,
@@ -206,7 +206,7 @@ def test_flashinfer_cutlass_moe_fp8_no_graph(
td = TestData.make_moe_tensors_8bit(m, k, n, e, reorder=False)
score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
topk_weights, topk_ids = FusedMoE.select_experts(
topk_weights, topk_ids, _ = FusedMoE.select_experts(
hidden_states=td.hidden_states,
router_logits=score,
use_grouped_topk=False,
+6
View File
@@ -273,6 +273,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
is_available_online=False),
"Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
is_available_online=False),
"LongcatFlashForCausalLM": _HfExamplesInfo
("meituan-longcat/LongCat-Flash-Chat", trust_remote_code=True),
"MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
"Mamba2ForCausalLM": _HfExamplesInfo("mistralai/Mamba-Codestral-7B-v0.1",
min_transformers_version="4.55.3",
@@ -639,6 +641,10 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
speculative_model="zai-org/GLM-4.5",
min_transformers_version="4.54",
is_available_online=False),
"LongCatFlashMTPModel": _HfExamplesInfo(
"meituan-longcat/LongCat-Flash-Chat",
trust_remote_code=True,
speculative_model="meituan-longcat/LongCat-Flash-Chat"),
"MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
trust_remote_code=True,
speculative_model="XiaomiMiMo/MiMo-7B-RL"),
+9 -3
View File
@@ -428,9 +428,8 @@ def dummy_hf_overrides(
num_hidden_layers = (3 if model_arch
== "Gemma3nForConditionalGeneration" else 1)
text_config.update({
update_dict = {
"num_layers": num_layers,
"num_hidden_layers": num_hidden_layers,
"num_experts": num_experts,
"num_experts_per_tok": 2,
"num_local_experts": num_experts,
@@ -440,7 +439,14 @@ def dummy_hf_overrides(
"n_routed_experts": num_experts,
# For Gemma-3n
"num_kv_shared_layers": 1,
})
}
# Update num_hidden_layers for non-Longcat architectures
if model_arch != "LongcatFlashForCausalLM" \
and model_arch != "LongCatFlashMTPModel":
update_dict["num_hidden_layers"] = num_hidden_layers
text_config.update(update_dict)
if hasattr(hf_config, "vision_config"):
hf_config.vision_config.update({
+1 -1
View File
@@ -96,7 +96,7 @@ def test_routing_strategy_integration(monkeypatch, device):
envs.environment_variables[env_name] = lambda s=strategy: s
# Test the select_experts method
topk_weights, topk_ids = FusedMoE.select_experts(
topk_weights, topk_ids, _ = FusedMoE.select_experts(
hidden_states=hidden_states,
router_logits=router_logits,
top_k=top_k,