Text Generation
Transformers
Safetensors
English
lfm2_moe
diffusion
block-diffusion
lfm2
Mixture of Experts
code
bf16
lora-merged
experimental
research-preview
conversational
Instructions to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="summerAIO/diffusion-lfm2.5-24b-a2b-blocktune") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("summerAIO/diffusion-lfm2.5-24b-a2b-blocktune") model = AutoModelForCausalLM.from_pretrained("summerAIO/diffusion-lfm2.5-24b-a2b-blocktune", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/summerAIO/diffusion-lfm2.5-24b-a2b-blocktune
- SGLang
How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with Docker Model Runner:
docker model run hf.co/summerAIO/diffusion-lfm2.5-24b-a2b-blocktune
File size: 9,912 Bytes
3605d88 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 | """Batched, answer-blind candidate selection extracted from measured code."""
import ast
from collections import Counter
import re
import time
import numpy as np
import torch
import diffusion_runtime as base
def forbidden_tokens(context, window, ngram, max_count, eos):
"""Constraints only inspect resolved left context; EOS is always permitted."""
context = context[-window:] if window else []
banned = {token for token, count in Counter(context).items() if max_count and count >= max_count}
if ngram and len(context) >= ngram - 1:
suffix = tuple(context[-(ngram - 1):]) if ngram > 1 else ()
for i in range(len(context) - ngram + 1):
if tuple(context[i:i + ngram - 1]) == suffix:
banned.add(context[i + ngram - 1])
return banned - set(eos)
@torch.inference_mode()
def sample_pool(sampler, prompt, cfg, length, seed):
"""Batched candidate generation, with CPU rejection resampling included in timing."""
model = sampler.model
eos_raw = model.config.eos_token_id
eos = set(eos_raw if isinstance(eos_raw,list) else [eos_raw])
eos_one = min(eos)
pad = model.config.pad_token_id
p = prompt.shape[1]
count = cfg["best_of"]
wave_size = min(count, max(1,4096//(p+length)))
rng = np.random.default_rng(seed)
prefix_list = prompt[0].tolist()
outputs, forwards = [], 0
for start in range(0,count,wave_size):
base.check_time(30)
batch = min(wave_size,count-start)
prefix = prompt.expand(batch,-1)
canvas = torch.full((batch,length),pad,device="cuda",dtype=torch.long)
mask = torch.ones_like(canvas,dtype=torch.bool)
cpu_ids = np.full((batch,length),-1,dtype=np.int64)
cpu_lp = np.zeros((batch,length),dtype=np.float64)
steps = min(cfg["steps"],length)
for step in range(steps):
base.check_time(30)
ids = torch.cat((prefix,canvas),1)
fullmask = torch.cat((torch.zeros_like(prefix,dtype=torch.bool),mask),1)
h = sampler.reference(ids,fullmask,p)[:,p:][mask]
logits = model.lm_head(h).float()
logits[:,pad] = -torch.inf
k = min(cfg["top_k"],logits.shape[-1]-1)
values, tokens = logits.topk(k,dim=-1)
# EOS is a genuine logit candidate even if outside top-k; never synthesize it as a hard-ban violation.
eos_values = logits[:,eos_one:eos_one+1]
eos_tokens = torch.full_like(tokens[:,:1],eos_one)
duplicate = tokens.eq(eos_one).any(-1,keepdim=True)
eos_values = eos_values.masked_fill(duplicate,-torch.inf)
values = torch.cat((values,eos_values),-1)
tokens = torch.cat((tokens,eos_tokens),-1)
normalizer = torch.logsumexp(logits/.9,-1).cpu().numpy()
raw_values = values.cpu().numpy()
token_values = tokens.cpu().numpy()
del logits, values, tokens, h
forwards += 1
remaining_after = length*(steps-step-1)//steps
row_offset = 0
for row in range(batch):
positions = np.flatnonzero(cpu_ids[row]<0)
commit = len(positions)-remaining_after
# Confidence selects slots; ascending positional commits enforce same-step bans.
maximum = raw_values[row_offset:row_offset+len(positions),0]/.9-normalizer[row_offset:row_offset+len(positions)]
chosen = np.argsort(maximum)[-commit:]
for local in sorted(chosen.tolist(),key=lambda x:int(positions[x])):
pos = int(positions[local])
idx = row_offset+local
left = pos-1
while left>=0 and cpu_ids[row,left]>=0:
left -= 1
context = cpu_ids[row,left+1:pos].tolist()
if left<0:
context = prefix_list+context
window_context = context[-cfg["window"]:] if cfg["window"] else []
banned = forbidden_tokens(context,cfg["window"],cfg["ngram"],cfg["max_repeat"],eos)
scores = raw_values[idx].astype(np.float64).copy()
if cfg["penalty"] != 1:
counts = Counter(window_context)
for j,token in enumerate(token_values[idx]):
if int(token) in counts and int(token) not in eos:
penalty = cfg["penalty"]**min(counts[int(token)],3)
scores[j] = scores[j]/penalty if scores[j]>=0 else scores[j]*penalty
scores /= .9
for j,token in enumerate(token_values[idx]):
if int(token) in banned:
scores[j] = -np.inf
assert np.isfinite(scores).any(), "EOS must remain selectable"
probability = np.exp(scores-np.max(scores))
probability /= probability.sum()
choice = int(rng.choice(len(scores),p=probability))
token = int(token_values[idx,choice])
assert token not in banned
cpu_ids[row,pos] = token
cpu_lp[row,pos] = float(raw_values[idx,choice]/.9-normalizer[idx])
row_offset += len(positions)
known = cpu_ids>=0
canvas = torch.as_tensor(np.where(known,cpu_ids,pad),device="cuda",dtype=torch.long)
mask = torch.as_tensor(~known,device="cuda")
for row in range(batch):
sequence = cpu_ids[row].tolist()
stop = next((i for i,token in enumerate(sequence) if token in eos),len(sequence))
assert all(t>=0 for t in sequence[:stop])
outputs.append({"ids":sequence[:stop],"ended":stop<len(sequence),
"mean_logprob":float(cpu_lp[row,:max(1,stop)].mean())})
return outputs,forwards,wave_size
def code_candidate(text, repair):
blocks = re.findall(r"```(?:python)?\s*\n(.*?)```",text,flags=re.S)
source = blocks[0] if blocks else text.strip()
choices = [source]
if repair:
lines = source.splitlines()
choices += ["\n".join(lines[:i]) for i in range(len(lines)-1,max(0,len(lines)-17),-1)]
for candidate in choices:
try:
tree = ast.parse(candidate)
definitions = [n for n in tree.body if isinstance(n,ast.FunctionDef)]
forbidden = any(isinstance(n,(ast.Import,ast.ImportFrom,ast.Attribute,ast.ClassDef)) for n in ast.walk(tree))
if definitions and not forbidden:
return candidate, [n.name for n in definitions], candidate!=source
except (SyntaxError,ValueError):
pass
return source,[],False
def rank_candidate(record, tokenizer, prompt_text, repair):
"""No answer or tests argument: selector cannot access evaluation truth."""
text = tokenizer.decode(record["ids"],skip_special_tokens=True).strip()
match = re.search(r"defining\s+([A-Za-z_]\w*)",prompt_text)
score = record["mean_logprob"]
changed = False
if match:
text,names,changed = code_candidate(text,repair)
score += 100 if match.group(1) in names else -100
score += 10 if names else 0
elif "only" in prompt_text.lower():
words = text.split()
score += 20 if 1<=len(words)<=4 and len(text)<60 else -20
health = base.health(record["ids"])
score -= max(0,health["max_run"]-3)*.5
score -= (1-health["distinct4"])*5
if not text:
score -= 1000
return score,text,changed
def generate_selected(sampler, tokenizer, prompt_text, cfg, seed, long=False):
initial = base.encode(tokenizer,prompt_text)
torch.cuda.synchronize()
start = time.perf_counter()
generated, gross, calls, chunks, repaired, max_batch = [],0,0,0,False,0
selected_text = ""
target = max(1024,cfg["block"]) if long else min(256,cfg["block"])
while len(generated)<target:
n = min(cfg["block"],target-len(generated))
history = initial.new_tensor([generated[-256:]])
prompt = torch.cat((initial,history),1)
pool,forwards,wave = sample_pool(sampler,prompt,cfg,n,seed+chunks*1009)
ranked = [(rank_candidate(row,tokenizer,prompt_text,cfg["repair"] and not long),index) for index,row in enumerate(pool)]
ranking,index = max(ranked,key=lambda item:item[0][0])
selected = pool[index]
selected_text,changed = ranking[1],ranking[2]
repaired |= changed
gross += sum(len(row["ids"]) for row in pool)
calls += forwards
max_batch = max(max_batch,wave)
generated.extend(selected["ids"])
chunks += 1
if not long or selected["ended"]:
break
if long:
selected_text = tokenizer.decode(generated,skip_special_tokens=True)
delivered = len(generated)
else:
# Post-processed output tokens, not discarded syntax-repair suffixes.
delivered = len(tokenizer(selected_text,add_special_tokens=False)["input_ids"])
torch.cuda.synchronize()
elapsed = time.perf_counter()-start
return selected_text,{"selected_tokens":delivered,"gross_candidate_tokens":gross,"seconds":elapsed,
"selected_tokens_per_second":delivered/elapsed,"gross_tokens_per_second":gross/elapsed,
"best_of":cfg["best_of"],"max_parallel_candidates":max_batch,"forward_calls":calls,
"chunks":chunks,"syntax_repaired":repaired,**base.health(generated)}
def cfg(block,steps=1,best_of=1,window=0,ngram=0,max_repeat=0,penalty=1.,top_k=50,repair=False):
return dict(block=block,steps=steps,best_of=best_of,window=window,ngram=ngram,max_repeat=max_repeat,
penalty=penalty,top_k=top_k,repair=repair,temperature=.9)
|