GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Crusadersk
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Crusadersk
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
#### number, with no
format bonus or zero-variance-group filtering.This is a single-GPU sequential mechanism screen. Projected overlap is not measured overlap. The portable evidence does not retain full vocabulary probabilities, so it cannot independently replay every verifier calculation. Parallel scoring and cropped-KV continuation are not established to share the same trajectory law as fresh incremental generation. The same rollout law and learner-update kernel remain unverified. This checkpoint has no independent downstream evaluation, safety qualification, or production-readiness claim.
| Arm | After-update time / fresh | Draft-inclusive time / fresh | Retained / drafted tokens |
|---|---|---|---|
| Fresh target | 1.0000 | 1.0000 | Not applicable |
| Tokenwise | 0.4289 | 1.4278 | 7,574 / 13,512 (56.05%) |
| Block verification | 0.2827 | 1.2816 | 10,167 / 13,512 (75.24%) |
| Suffix cache | 0.8533 | 1.8522 | Not comparable: repeated proposals |
These are ratios of summed measured times across all 64 future slots;
below 1 favors reuse. With S = serial fresh-target time, V = correction/cache
time, and D = draft generation charged once per draw, the two columns are
V/S and (D+V)/S. The shared two-draft cache bank is charged once for its two
outputs. Retention counts are realized outcomes, not probabilities of complete
trajectory acceptance; repeated cache proposals do not share that denominator.
D includes original-law storage during generation. These component ratios
exclude native training collection, optimizer updates, setup, checkpoint/snapshot
overhead and diagnostic work; full before-hook costs and ideal-overlap projections
are separately reported in the analysis. Output lengths and cap counts differ by
arm. The fresh comparator is serial Actor generation, not optimized batched
serving. High retention or a favorable component ratio establishes neither equal
end-to-end sampling laws nor training-throughput improvement.
The complete descriptive analysis is evidence/analysis.json.
portable_experiment.zip contains every original non-weight run artifact,
including the unchanged original hash manifest and checkpoint configurations.
All nine original checkpoints are also available under checkpoints/step-0
through checkpoints/step-8. The root weights equal step 8 byte-for-byte;
only the root adapter configuration receives the pinned base revision.
publication_manifest.json records original/published hashes and omissions.
Source: GitHub PR #13, frozen protocol/source commit. The original configurations are historical evidence: when loading an archived checkpoint, explicitly load the pinned base shown below.
The example pins the initial published weights and inference settings. The reference environment used Transformers 5.10.1 and PEFT 0.19.1.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig adapter_repo = "Crusadersk/chimera-rollout-futures-qwen2.5-7b-grpo-screen-20261012"adapter_revision = "7cfa299a9df91a456ddf4e3d180191432cd9303b"base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen2.5-7B-Instruct", revision="a09a35458c702b33eeacc393d103063234e8bc28", dtype=torch.bfloat16, device_map="auto", trust_remote_code=False,)model = PeftModel.from_pretrained(base, adapter_repo, revision=adapter_revision).eval()tokenizer = AutoTokenizer.from_pretrained(adapter_repo, revision=adapter_revision)model.generation_config = GenerationConfig.from_pretrained(adapter_repo, revision=adapter_revision)messages = [{"role": "user", "content": "What is 17 + 25? Solve the problem and finish with #### followed by the numerical answer."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True,).to(model.device)with torch.inference_mode(): output = model.generate(**inputs, generation_config=model.generation_config)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))The example is ordinary adapter inference. It does not implement or validate Rollout Futures speculative verification.
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig adapter_repo = "Crusadersk/chimera-rollout-futures-qwen2.5-7b-grpo-screen-20261012"adapter_revision = "7cfa299a9df91a456ddf4e3d180191432cd9303b"base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen2.5-7B-Instruct", revision="a09a35458c702b33eeacc393d103063234e8bc28", dtype=torch.bfloat16, device_map="auto", trust_remote_code=False,)model = PeftModel.from_pretrained(base, adapter_repo, revision=adapter_revision).eval()tokenizer = AutoTokenizer.from_pretrained(adapter_repo, revision=adapter_revision)model.generation_config = GenerationConfig.from_pretrained(adapter_repo, revision=adapter_revision)messages = [{"role": "user", "content": "What is 17 + 25? Solve the problem and finish with #### followed by the numerical answer."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True,).to(model.device)with torch.inference_mode(): output = model.generate(**inputs, generation_config=model.generation_config)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))