---
library_name: transformers
base_model:
- Qwen/Qwen3.8-Flash-Next-FP8
---
This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from [Qwen/Qwen3.8-Flash-Next-FP8](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8).
Structural comparison with `Qwen/Qwen3.8-Flash-Next-FP8`:
| Feature | Source | Tiny |
|---|---|---|
| Architecture / modalities | Qwen4Exp; text, image, video | Same |
| Vocabulary / tied embeddings | 248,320 / false | Same |
| Backbone layers / hidden size | 48 / 2,560 | 8 / 32 |
| Attention Q/KV heads; head dim | 24/2; 256 | 12/1; 32 |
| Attention cycle | 3 linear + 1 QSA | Same; two cycles |
| Linear key/value heads; head dim | 16/48; 128 | 4/12; 32 |
| Routed experts / top-k | 512 / 10 | 32 / 10 |
| Expert/shared intermediate size | 640 / 640 | 128 / 128 |
| Hyper-Connections / mixing rank | 4 / 320 | 4 / 4 |
| QSA heads / dimension / token budget / compression | 4 / 128 / 2,048 / 4 | Same |
| PLE layer / n-gram orders / heads per order | 2 / 2,3 / 8 | Same |
| PLE rows / embedding columns / dtype | 320,001,536 / 160 / FP8 | 9,216 / 1 / FP8 |
| PLE hash bucket base | 20,000,000 | 512; official prime layout |
| PLE convolution / serialized table shards | 4 / 128 | Same |
| MTP | 1 QSA layer; HC fusion; shared token embeddings | Same |
| Vision depth / hidden size / output size | 27 / 1,152 / 2,560 | 2 / 64 / 32 |
| Vision patch / temporal patch / merge | 16 / 2 / 2 | Same |
PLE retains the source n-gram orders, head count, hash/prime-layout algorithm,
projections and convolution with reduced auxiliary hash capacity. The main token
vocabulary and IDs are unchanged.
The two untied BF16 token embedding/head matrices at width 32 alone occupy
31,784,960 bytes, exceeding the 30 MB target. This native block-FP8 backend
requires width 32; the complete checkpoint is 37,655,018 bytes. Routed experts
retain FP8 with 32x32 blocks and BF16 scales; PLE retains FP8 and one BF16 scale.
Tokenizer and processor assets are copied unchanged from the pinned source.
| File path | Size |
|------|------|
| model.safetensors | 37.7MB |
### Example usage:
- SGLang (✅ verified)
- Run on both two-GPU nodes with NODE_RANK set to 0 or 1, the same
DIST_INIT_ADDR, and an available local NCCL_PORT. Enables TP=4, NEXTN/MTP
and CUDA Graph.
- Verified on four NVIDIA H20 GPUs across two nodes for text, QSA long
prompts, image-plus-text and video-plus-text; see Test environment.
```bash
model_id=tiny-random/qwen3.8-flash-next-fp8
node_rank=${NODE_RANK:-0}
dist_init_addr=${DIST_INIT_ADDR}
nccl_port=${NCCL_PORT}
sglang serve \
--model-path $model_id \
--tp-size 4 \
--nnodes 2 \
--node-rank $node_rank \
--dist-init-addr $dist_init_addr \
--nccl-port $nccl_port \
--fp8-gemm-backend triton \
--moe-runner-backend triton \
--speculative-algorithm NEXTN \
--speculative-num-steps 3 \
--speculative-eagle-topk 1 \
--speculative-num-draft-tokens 4 \
--context-length 4096 \
--max-total-tokens 4096 \
--chunked-prefill-size 512 \
--mem-fraction-static 0.4 \
--cuda-graph-config '{"decode": {"bs": [1, 2, 4, 8]}}' \
--disable-custom-all-reduce
```
- Transformers (✅ verified)
```python
import numpy as np
import torch
from PIL import Image
from transformers import AutoProcessor, Qwen4ExpForConditionalGeneration
from transformers.video_utils import VideoMetadata
model_id = "tiny-random/qwen3.8-flash-next-fp8"
model = Qwen4ExpForConditionalGeneration.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="cuda", attn_implementation="sdpa",
)
processor = AutoProcessor.from_pretrained(model_id)
image = Image.new("RGB", (64, 64), (220, 30, 30))
# Four predecoded RGB frames, with timestamps supplied through video_metadata.
video = np.stack([np.asarray(Image.new("RGB", (64, 64), color)) for color in
[(220, 30, 30), (30, 220, 30), (30, 30, 220), (220, 220, 30)]])
messages = [{"role": "user", "content": [
{"type": "image", "image": image},
{"type": "video", "video": video},
{"type": "text", "text": "Describe the image and the video."},
]}]
inputs = processor.apply_chat_template(
messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",
do_sample_frames=False,
video_metadata=[VideoMetadata(total_num_frames=4, fps=2, frames_indices=[0, 1, 2, 3])],
).to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=10)
print(processor.decode(output[0, inputs['input_ids'].shape[1]:], skip_special_tokens=True))
```
### Codes to create this repo (✅ verified):
Click to expand
```python
import json
from copy import deepcopy
from pathlib import Path
from shutil import copyfile
import torch
from huggingface_hub import hf_hub_download, list_repo_files
from safetensors.torch import save_file
from transformers import AutoConfig, AutoModelForImageTextToText, set_seed
from transformers.models.qwen4_exp.modeling_qwen4_exp import Qwen4ExpTextRMSNorm
source_model_id = "Qwen/Qwen3.8-Flash-Next-FP8"
save_folder = "/tmp/tiny-random/qwen38-flash-next-fp8"
source_revision = '236dfdf285828023ca3bcd3f37366c58a3469b13'
set_seed(42)
folder = Path(save_folder)
folder.mkdir(parents=True, exist_ok=True)
raw = json.loads(Path(hf_hub_download(source_model_id, 'config.json', revision=source_revision)).read_text())
quantization = raw.pop('quantization_config')
quantization['weight_block_size'] = [32, 32]
raw['text_config'].update(
hidden_size=32, ple_embed_dim=16, ngram_vocab_size_base=512, hc_lowrank=4, num_hidden_layers=8,
layer_types=(["linear_attention"] * 3 + ["full_attention"]) * 2,
num_attention_heads=12, num_key_value_heads=1, head_dim=32,
linear_num_key_heads=4, linear_num_value_heads=12,
linear_key_head_dim=32, linear_value_head_dim=32,
num_experts=32, moe_intermediate_size=128, shared_expert_intermediate_size=128,
)
raw['text_config']['rope_parameters']['mrope_section'] = [2, 1, 1]
raw['vision_config'].update(depth=2, hidden_size=64, intermediate_size=128, num_heads=4, out_hidden_size=32)
(folder / 'config.json').write_text(json.dumps(raw, indent=2))
config = AutoConfig.from_pretrained(folder)
model = AutoModelForImageTextToText.from_config(config, dtype=torch.bfloat16)
backbone = model.model.language_model
text = backbone.config
width = text.hidden_size
model.mtp = torch.nn.ModuleDict({
"fc_embedding": torch.nn.Linear(width, width, bias=False, dtype=torch.bfloat16),
"fc_hidden": torch.nn.Linear(width, width, bias=False, dtype=torch.bfloat16),
"pre_fc_norm_embedding": Qwen4ExpTextRMSNorm(width, eps=text.rms_norm_eps).to(torch.bfloat16),
"pre_fc_norm_hidden": Qwen4ExpTextRMSNorm(width * text.hc_count, eps=text.rms_norm_eps).to(torch.bfloat16),
"layers": torch.nn.ModuleList([deepcopy(backbone.layers[-1])]),
"hyper_connection_mixer": deepcopy(backbone.hyper_connection_mixer),
})
with torch.no_grad():
for name, param in sorted(model.named_parameters()):
if name.endswith("A_log"):
param.zero_()
elif name.endswith("dt_bias"):
param.fill_(1)
elif "norm" in name and name.endswith("weight"):
param.fill_(1 if "visual" in name or ".linear_attn.norm." in name else 0)
elif name.endswith("bias"):
param.zero_()
else:
torch.nn.init.normal_(param, mean=0, std=0.02)
state = {name: value.detach().cpu().contiguous() for name, value in model.state_dict().items()}
# Prune source exclusions absent from this architecture before FP8 loading.
names = set(dict(model.named_modules())) | set(dict(model.named_parameters()))
exclusions = [name for name in quantization["modules_to_not_convert"] if name in names]
quantization["modules_to_not_convert"] = exclusions
def pack(name, weight):
n, k = weight.shape
bn, bk = quantization["weight_block_size"]
blocks = weight.float().reshape(n // bn, bn, k // bk, bk)
scales = (blocks.abs().amax(dim=(1, 3)) / 448).clamp_min(1e-12).to(torch.bfloat16)
packed = (blocks / scales.float()[:, None, :, None]).clamp(-448, 448)
state[name] = packed.reshape(n, k).to(torch.float8_e4m3fn).contiguous()
state[name.removesuffix("weight") + "weight_scale_inv"] = scales.contiguous()
for name, module in model.named_modules():
excluded = any(name == prefix or name.startswith(prefix + ".") for prefix in exclusions)
if isinstance(module, torch.nn.Linear) and not excluded:
pack(name + ".weight", state[name + ".weight"])
elif isinstance(module, torch.nn.Embedding) and not excluded and any(
name.endswith(pattern) for pattern in quantization.get("modules_to_convert", ())
):
table = state.pop(name + ".weight")
scale = torch.tensor([0.001], dtype=torch.bfloat16)
packed = torch.empty_like(table, dtype=torch.float8_e4m3fn)
for start in range(0, table.shape[0], 1_000_000):
packed[start:start + 1_000_000] = (
table[start:start + 1_000_000].float() / scale.float()
).clamp(-448, 448).to(torch.float8_e4m3fn)
state[name + ".weight"] = packed
state[name + ".weight_scale"] = scale
# Public FP8 checkpoints store three projections per routed expert.
for name in list(state):
if not name.endswith(".experts.gate_up_proj"):
continue
prefix = name.removesuffix(".gate_up_proj")
up = state.pop(name)
down = state.pop(prefix + ".down_proj")
for expert in range(up.shape[0]):
gate, value = up[expert].chunk(2, dim=0)
for projection, weight in (("gate_proj", gate), ("up_proj", value), ("down_proj", down[expert])):
pack(f"{prefix}.{expert}.{projection}.weight", weight)
# The public checkpoint stores PLE embeddings in shard_N tensors.
for name in list(state):
if not name.endswith(".ple.ple_embedding.ngram_embedding.weight"):
continue
table = state.pop(name)
parts = text.split_ngram_parts
rows = (table.shape[0] + parts - 1) // parts
prefix = name.removesuffix(".weight")
for part in range(parts):
state[f"{prefix}.shard_{part}.weight"] = table[part * rows:(part + 1) * rows].contiguous()
save_file(state, folder / "model.safetensors", metadata={"format": "pt"})
raw['quantization_config'] = quantization
(folder / 'config.json').write_text(json.dumps(raw, indent=2))
# Preserve tokenizer, processor and generation assets exactly as published.
assets = {
"tokenizer.json", "tokenizer_config.json", "special_tokens_map.json",
"vocab.json", "merges.txt", "added_tokens.json", "chat_template.jinja",
"chat_template.json", "preprocessor_config.json", "processor_config.json",
"video_preprocessor_config.json", "generation_config.json",
}
for name in sorted(set(list_repo_files(source_model_id, revision=source_revision)) & assets):
copyfile(hf_hub_download(source_model_id, name, revision=source_revision), folder / name)
```
Click to expand
```text
Qwen4ExpForConditionalGeneration(
(model): Qwen4ExpModel(
(visual): Qwen4ExpVisionModel(
(patch_embed): Qwen4ExpVisionPatchEmbed(
(proj): Conv3d(3, 64, kernel_size=(2, 16, 16), stride=(2, 16, 16))
)
(pos_embed): Embedding(2304, 64)
(rotary_pos_emb): Qwen4ExpVisionRotaryEmbedding()
(blocks): ModuleList(
(0-1): 2 x Qwen4ExpVisionBlock(
(norm1): LayerNorm((64,), eps=1e-06, elementwise_affine=True, bias=True)
(norm2): LayerNorm((64,), eps=1e-06, elementwise_affine=True, bias=True)
(attn): Qwen4ExpVisionAttention(
(qkv): Linear(in_features=64, out_features=192, bias=True)
(proj): Linear(in_features=64, out_features=64, bias=True)
)
(mlp): Qwen4ExpVisionMLP(
(linear_fc1): Linear(in_features=64, out_features=128, bias=True)
(linear_fc2): Linear(in_features=128, out_features=64, bias=True)
(act_fn): GELUTanh()
)
)
)
(merger): Qwen4ExpVisionPatchMerger(
(norm): LayerNorm((64,), eps=1e-06, elementwise_affine=True, bias=True)
(linear_fc1): Linear(in_features=256, out_features=256, bias=True)
(act_fn): GELU(approximate='none')
(linear_fc2): Linear(in_features=256, out_features=32, bias=True)
)
)
(language_model): Qwen4ExpTextModel(
(embed_tokens): Embedding(248320, 32)
(layers): ModuleList(
(0): Qwen4ExpTextDecoderLayer(
(linear_attn): Qwen4ExpTextGatedDeltaNet(
(conv1d): Conv1d(640, 640, kernel_size=(4,), stride=(1,), padding=(3,), groups=640, bias=False)
(norm): Qwen4ExpTextRMSNormGated()
(out_proj): Linear(in_features=384, out_features=32, bias=False)
(in_proj_qkv): Linear(in_features=32, out_features=640, bias=False)
(in_proj_z): Linear(in_features=32, out_features=384, bias=False)
(in_proj_b): Linear(in_features=32, out_features=12, bias=False)
(in_proj_a): Linear(in_features=32, out_features=12, bias=False)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
(1): Qwen4ExpTextDecoderLayer(
(linear_attn): Qwen4ExpTextGatedDeltaNet(
(conv1d): Conv1d(640, 640, kernel_size=(4,), stride=(1,), padding=(3,), groups=640, bias=False)
(norm): Qwen4ExpTextRMSNormGated()
(out_proj): Linear(in_features=384, out_features=32, bias=False)
(in_proj_qkv): Linear(in_features=32, out_features=640, bias=False)
(in_proj_z): Linear(in_features=32, out_features=384, bias=False)
(in_proj_b): Linear(in_features=32, out_features=12, bias=False)
(in_proj_a): Linear(in_features=32, out_features=12, bias=False)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(ple): Qwen4ExpTextPLELayer(
(ple_embedding): Qwen4ExpTextNGramEmbedding(
(ngram_embedding): Embedding(9216, 1)
)
(key_proj): Linear(in_features=16, out_features=128, bias=False)
(value_proj): Linear(in_features=16, out_features=32, bias=False)
(norm_key): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(norm_query): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(norm_conv): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(conv1d): Conv1d(128, 128, kernel_size=(4,), stride=(1,), dilation=(3,), groups=128, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
(2): Qwen4ExpTextDecoderLayer(
(linear_attn): Qwen4ExpTextGatedDeltaNet(
(conv1d): Conv1d(640, 640, kernel_size=(4,), stride=(1,), padding=(3,), groups=640, bias=False)
(norm): Qwen4ExpTextRMSNormGated()
(out_proj): Linear(in_features=384, out_features=32, bias=False)
(in_proj_qkv): Linear(in_features=32, out_features=640, bias=False)
(in_proj_z): Linear(in_features=32, out_features=384, bias=False)
(in_proj_b): Linear(in_features=32, out_features=12, bias=False)
(in_proj_a): Linear(in_features=32, out_features=12, bias=False)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
(3): Qwen4ExpTextDecoderLayer(
(self_attn): Qwen4ExpTextAttention(
(q_proj): Linear(in_features=32, out_features=768, bias=False)
(k_proj): Linear(in_features=32, out_features=32, bias=False)
(v_proj): Linear(in_features=32, out_features=32, bias=False)
(o_proj): Linear(in_features=384, out_features=32, bias=False)
(q_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(k_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(indexer): Qwen4ExpTextQSAIndexer(
(index_qk_proj): Linear(in_features=32, out_features=640, bias=False)
(q_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(k_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
(4-6): 3 x Qwen4ExpTextDecoderLayer(
(linear_attn): Qwen4ExpTextGatedDeltaNet(
(conv1d): Conv1d(640, 640, kernel_size=(4,), stride=(1,), padding=(3,), groups=640, bias=False)
(norm): Qwen4ExpTextRMSNormGated()
(out_proj): Linear(in_features=384, out_features=32, bias=False)
(in_proj_qkv): Linear(in_features=32, out_features=640, bias=False)
(in_proj_z): Linear(in_features=32, out_features=384, bias=False)
(in_proj_b): Linear(in_features=32, out_features=12, bias=False)
(in_proj_a): Linear(in_features=32, out_features=12, bias=False)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
(7): Qwen4ExpTextDecoderLayer(
(self_attn): Qwen4ExpTextAttention(
(q_proj): Linear(in_features=32, out_features=768, bias=False)
(k_proj): Linear(in_features=32, out_features=32, bias=False)
(v_proj): Linear(in_features=32, out_features=32, bias=False)
(o_proj): Linear(in_features=384, out_features=32, bias=False)
(q_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(k_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(indexer): Qwen4ExpTextQSAIndexer(
(index_qk_proj): Linear(in_features=32, out_features=640, bias=False)
(q_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(k_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
)
(rotary_emb): Qwen4ExpTextRotaryEmbedding()
(hyper_connection_mixer): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
)
)
)
(lm_head): Linear(in_features=32, out_features=248320, bias=False)
(mtp): ModuleDict(
(fc_embedding): Linear(in_features=32, out_features=32, bias=False)
(fc_hidden): Linear(in_features=32, out_features=32, bias=False)
(pre_fc_norm_embedding): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(pre_fc_norm_hidden): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(layers): ModuleList(
(0): Qwen4ExpTextDecoderLayer(
(self_attn): Qwen4ExpTextAttention(
(q_proj): Linear(in_features=32, out_features=768, bias=False)
(k_proj): Linear(in_features=32, out_features=32, bias=False)
(v_proj): Linear(in_features=32, out_features=32, bias=False)
(o_proj): Linear(in_features=384, out_features=32, bias=False)
(q_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(k_norm): Qwen4ExpTextRMSNorm((32,), eps=1e-06)
(indexer): Qwen4ExpTextQSAIndexer(
(index_qk_proj): Linear(in_features=32, out_features=640, bias=False)
(q_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(k_layernorm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
)
)
(mlp): Qwen4ExpTextSparseMoeBlock(
(gate): Qwen4ExpTextTopKRouter()
(experts): Qwen4ExpTextExperts(
(act_fn): SiLUActivation()
)
(shared_expert): Qwen4ExpTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLUActivation()
)
(shared_expert_gate): Linear(in_features=32, out_features=1, bias=False)
)
(attn_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
(mlp_hyper_connection): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
(block_inject_weight): Linear(in_features=128, out_features=4, bias=False)
)
)
)
(hyper_connection_mixer): Qwen4ExpTextGatedResidual(
(hc_norm): Qwen4ExpTextRMSNorm((128,), eps=1e-06)
(input_mix_weight_down): Linear(in_features=128, out_features=4, bias=False)
(input_mix_weight_up): Linear(in_features=4, out_features=128, bias=False)
)
)
)
```