Publish openpeerllm-demo

README.md

@@ -1,13 +1,31 @@
----
-title: OpenPeerLLM Demo
-emoji: 🌍
-colorFrom: blue
-colorTo: indigo
-sdk: gradio
-sdk_version: 6.30.0
-python_version: '3.12'
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: OpenPeerLLM Demo
+emoji: 🕸️
+colorFrom: blue
+colorTo: green
+sdk: gradio
+sdk_version: 5.49.1
+app_file: app.py
+pinned: false
+license: mit
+models:
+  - OpenPeerAI/OpenPeerLLM
+short_description: Generate and score text with OpenPeerLLM on ZeroGPU
+---
+
+# OpenPeerLLM Demo
+
+Text generation and perplexity scoring with
+[OpenPeerAI/OpenPeerLLM](https://huggingface.co/OpenPeerAI/OpenPeerLLM), a
+decentralised-training language model by Andrew Magdy Kamal Nassief
+(DOI [10.57967/hf/6469](https://doi.org/10.57967/hf/6469)).
+
+The space loads OpenPeerLLM's published `model.safetensors` (163M parameters)
+tensor for tensor into a reimplementation of its architecture that adds the
+causal attention mask the reference forward pass omits. OpenPeerLLM is the
+starting rung of the
+[My Name is Riemann](https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease)
+growth ladder.
+
+Maintained by OpenPeer AI. OpenPeerLLM is released under OPNL, MIT and
+CC BY 4.0.

app.py

@@ -0,0 +1,133 @@
+from __future__ import annotations
+import threading
+import torch
+import gradio as gr
+import spaces
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+from transformers import AutoTokenizer, TextIteratorStreamer
+from configuration_riemann import RiemannConfig
+from modeling_riemann import RiemannForCausalLM
+
+OPENPEER_REPO = "OpenPeerAI/OpenPeerLLM"
+
+tokenizer = AutoTokenizer.from_pretrained("gpt2")
+tokenizer.pad_token = tokenizer.eos_token
+config = RiemannConfig(
+    vocab_size=50257,
+    hidden_size=768,
+    num_hidden_layers=12,
+    num_attention_heads=12,
+    intermediate_size=3072,
+    max_position_embeddings=1024,
+)
+model = RiemannForCausalLM(config)
+weights_path = hf_hub_download(OPENPEER_REPO, "model.safetensors")
+state = load_file(weights_path)
+missing, unexpected = model.load_state_dict(state, strict=False)
+LOAD_REPORT = f"Loaded {len(state)} tensors from {OPENPEER_REPO}/model.safetensors. Missing: {len(missing)}. Unexpected: {len(unexpected)}."
+model.eval()
+N_PARAMS = sum(p.numel() for p in model.parameters())
+
+
+@spaces.GPU(duration=60)
+def generate(prompt, max_new_tokens, temperature, top_p, repetition_penalty):
+    model.to("cuda")
+    inputs = tokenizer(
+        prompt,
+        return_tensors="pt",
+        truncation=True,
+        max_length=1024 - int(max_new_tokens),
+    ).to("cuda")
+    streamer = TextIteratorStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+    kwargs = dict(
+        **inputs,
+        streamer=streamer,
+        max_new_tokens=int(max_new_tokens),
+        do_sample=temperature > 0,
+        temperature=max(float(temperature), 1e-3),
+        top_p=float(top_p),
+        repetition_penalty=float(repetition_penalty),
+        pad_token_id=tokenizer.eos_token_id,
+    )
+    thread = threading.Thread(target=model.generate, kwargs=kwargs)
+    thread.start()
+    out = ""
+    for piece in streamer:
+        out += piece
+        yield prompt + out
+    thread.join()
+
+
+@spaces.GPU(duration=30)
+def score(text):
+    model.to("cuda")
+    ids = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024).to(
+        "cuda"
+    )
+    with torch.no_grad():
+        out = model(input_ids=ids["input_ids"], labels=ids["input_ids"])
+    n = ids["input_ids"].shape[1]
+    return f"Tokens: {n}\nMean NLL: {out.loss.item():.4f}\nPerplexity: {torch.exp(out.loss).item():.2f}"
+
+
+DESCRIPTION = f"""
+# OpenPeerLLM Demo
+
+[OpenPeerAI/OpenPeerLLM](https://huggingface.co/{OPENPEER_REPO}) is a decentralised-training
+language model by Andrew Magdy Kamal Nassief (DOI [10.57967/hf/6469](https://doi.org/10.57967/hf/6469)).
+This space loads its published `model.safetensors` ({N_PARAMS/1e6:.1f}M parameters, 12 layers,
+hidden size 768, GPT-2 vocabulary) into a faithful reimplementation of its architecture
+and runs it on ZeroGPU.
+
+{LOAD_REPORT}
+
+OpenPeerLLM is the starting rung of the
+[My Name is Riemann](https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease) growth ladder.
+The reimplementation here adds a causal attention mask, which the reference forward pass omits and
+which is required for coherent autoregressive generation.
+"""
+
+with gr.Blocks(title="OpenPeerLLM Demo") as demo:
+    gr.Markdown(DESCRIPTION)
+    with gr.Tab("Generate"):
+        prompt = gr.Textbox(
+            label="Prompt",
+            lines=4,
+            value="The theory of decentralised machine learning says that",
+        )
+        with gr.Row():
+            max_new = gr.Slider(8, 256, value=96, step=8, label="Max new tokens")
+            temp = gr.Slider(
+                0.0, 1.5, value=0.8, step=0.05, label="Temperature (0 = greedy)"
+            )
+            top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
+            rep = gr.Slider(1.0, 1.5, value=1.2, step=0.05, label="Repetition penalty")
+        out = gr.Textbox(label="Output", lines=10)
+        gr.Button("Generate", variant="primary").click(
+            generate, [prompt, max_new, temp, top_p, rep], out
+        )
+        gr.Examples(
+            [
+                ["Peer-to-peer networks allow"],
+                ["def fibonacci(n):"],
+                ["In 1859, Bernhard Riemann"],
+            ],
+            inputs=prompt,
+        )
+    with gr.Tab("Score text"):
+        txt = gr.Textbox(
+            label="Text to score",
+            lines=6,
+            value="Decentralized training distributes model updates across peers.",
+        )
+        res = gr.Textbox(label="Negative log-likelihood and perplexity", lines=3)
+        gr.Button("Score", variant="primary").click(score, txt, res)
+    gr.Markdown(
+        "OpenPeerLLM is released under OPNL, MIT and CC BY 4.0. "
+        "This demo is maintained by OpenPeer AI."
+    )
+
+demo.queue(max_size=16).launch()

configuration_riemann.py

@@ -0,0 +1,33 @@
+from __future__ import annotations
+from transformers import PretrainedConfig
+
+
+class RiemannConfig(PretrainedConfig):
+    model_type = "riemann_openpeer"
+
+    def __init__(
+        self,
+        vocab_size: int = 50257,
+        hidden_size: int = 768,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        intermediate_size: int = 3072,
+        max_position_embeddings: int = 1024,
+        layer_norm_eps: float = 1e-05,
+        hidden_dropout: float = 0.1,
+        attention_dropout: float = 0.1,
+        use_gradient_checkpointing: bool = False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.use_gradient_checkpointing = use_gradient_checkpointing
+        kwargs.setdefault("tie_word_embeddings", False)
+        super().__init__(**kwargs)

modeling_riemann.py

@@ -0,0 +1,244 @@
+from __future__ import annotations
+import math
+from typing import Optional
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from transformers import GenerationMixin, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from configuration_riemann import RiemannConfig
+
+
+class MultiHeadAttention(nn.Module):
+
+    def __init__(self, config: RiemannConfig):
+        super().__init__()
+        self.num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.num_heads
+        self.query = nn.Linear(config.hidden_size, config.hidden_size)
+        self.key = nn.Linear(config.hidden_size, config.hidden_size)
+        self.value = nn.Linear(config.hidden_size, config.hidden_size)
+        self.out = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+    def forward(
+        self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
+    ):
+        batch_size, seq_length = hidden_states.shape[:2]
+        query_states = self.query(hidden_states)
+        key_states = self.key(hidden_states)
+        value_states = self.value(hidden_states)
+        query_states = query_states.view(
+            batch_size, seq_length, self.num_heads, self.head_size
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            batch_size, seq_length, self.num_heads, self.head_size
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            batch_size, seq_length, self.num_heads, self.head_size
+        ).transpose(1, 2)
+        use_sdpa = hidden_states.device.type in ("cuda", "cpu")
+        if use_sdpa:
+            context_layer = F.scaled_dot_product_attention(
+                query_states,
+                key_states,
+                value_states,
+                attn_mask=attention_mask,
+                dropout_p=self.dropout.p if self.training else 0.0,
+                is_causal=attention_mask is None,
+            )
+        else:
+            attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))
+            attention_scores = attention_scores / math.sqrt(self.head_size)
+            if attention_mask is None:
+                attention_mask = torch.triu(
+                    torch.full(
+                        (seq_length, seq_length),
+                        torch.finfo(attention_scores.dtype).min,
+                        dtype=attention_scores.dtype,
+                        device=attention_scores.device,
+                    ),
+                    diagonal=1,
+                )
+            attention_scores = attention_scores + attention_mask
+            attention_probs = F.softmax(attention_scores, dim=-1)
+            attention_probs = self.dropout(attention_probs)
+            context_layer = torch.matmul(attention_probs, value_states)
+        context_layer = context_layer.transpose(1, 2).contiguous()
+        context_layer = context_layer.view(batch_size, seq_length, self.hidden_size)
+        context_layer = self.out(context_layer)
+        return context_layer
+
+
+class MLP(nn.Module):
+
+    def __init__(self, config: RiemannConfig):
+        super().__init__()
+        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.act = nn.GELU()
+        self.dropout = nn.Dropout(config.hidden_dropout)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense_h_to_4h(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.dense_4h_to_h(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class TransformerBlock(nn.Module):
+
+    def __init__(self, config: RiemannConfig):
+        super().__init__()
+        self.attention = MultiHeadAttention(config)
+        self.mlp = MLP(config)
+        self.input_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps
+        )
+        self.post_attention_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps
+        )
+        self.dropout = nn.Dropout(config.hidden_dropout)
+
+    def forward(
+        self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
+    ):
+        attention_layernorm_out = self.input_layernorm(hidden_states)
+        attention_output = self.attention(
+            attention_layernorm_out, attention_mask=attention_mask
+        )
+        attention_output = self.dropout(attention_output)
+        attention_output = attention_output + hidden_states
+        mlp_layernorm_out = self.post_attention_layernorm(attention_output)
+        mlp_output = self.mlp(mlp_layernorm_out)
+        layer_output = mlp_output + attention_output
+        return layer_output
+
+
+class RiemannPreTrainedModel(PreTrainedModel):
+    config_class = RiemannConfig
+    base_model_prefix = "riemann"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+class RiemannForCausalLM(RiemannPreTrainedModel, GenerationMixin):
+
+    def __init__(self, config: RiemannConfig):
+        super().__init__(config)
+        self.config = config
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.layers = nn.ModuleList(
+            [TransformerBlock(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.final_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps
+        )
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.word_embeddings = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, value):
+        self.lm_head = value
+
+    def _build_attention_mask(self, input_shape, attention_mask, dtype, device):
+        batch_size, seq_length = input_shape
+        causal = torch.triu(
+            torch.full(
+                (seq_length, seq_length),
+                torch.finfo(dtype).min,
+                dtype=dtype,
+                device=device,
+            ),
+            diagonal=1,
+        )
+        causal = causal.unsqueeze(0).unsqueeze(0)
+        if attention_mask is not None:
+            padding = attention_mask[:, None, None, :].to(dtype=dtype)
+            padding = (1.0 - padding) * torch.finfo(dtype).min
+            return causal + padding
+        return causal
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        **kwargs,
+    ):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("pass either input_ids or inputs_embeds, not both")
+        if input_ids is not None:
+            inputs_embeds = self.word_embeddings(input_ids)
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("need either input_ids or inputs_embeds")
+        position_ids = torch.arange(
+            seq_length, dtype=torch.long, device=inputs_embeds.device
+        )
+        position_ids = position_ids.unsqueeze(0).expand(batch_size, -1)
+        position_embeds = self.position_embeddings(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+        full_mask = (
+            None
+            if attention_mask is None
+            else self._build_attention_mask(
+                (batch_size, seq_length),
+                attention_mask,
+                hidden_states.dtype,
+                hidden_states.device,
+            )
+        )
+        for layer in self.layers:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    layer, hidden_states, full_mask, use_reentrant=False
+                )
+            else:
+                hidden_states = layer(hidden_states, full_mask)
+        hidden_states = self.final_layernorm(hidden_states)
+        logits = self.lm_head(hidden_states)
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss = F.cross_entropy(
+                shift_logits.view(-1, self.config.vocab_size),
+                shift_labels.view(-1),
+                ignore_index=-100,
+            )
+        return CausalLMOutputWithPast(
+            loss=loss, logits=logits, hidden_states=hidden_states
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        return {"input_ids": input_ids}

requirements.txt

@@ -0,0 +1,6 @@
+torch>=2.4
+transformers>=4.56,<5
+safetensors>=0.4
+gradio>=5.0
+spaces
+huggingface_hub>=0.24