start LLM app, tons of clean up required. target is 200 line ollama (#11068)

* start LLM app, tons of clean up required. target is 200 line ollama

* kind of works

* simpler

* add k/v cache

* with SYM=1, it loops

* no rope cache

* simpler

* more cleanups

* cleanups

* works

* argparse and comments

* from gguf

* generate is a function

* no copy from cpu

* fix max context pass in

* test

* improve test

* ai2_arc

* fix 8B, use less ram

* 136 lines

.github/workflows/test.yml

@@ -368,8 +368,8 @@ jobs:
         PYTHONPATH=. python extra/optimization/extract_dataset.py
         gzip -c /tmp/sops > extra/datasets/sops.gz
         DEBUG=1 MIN_ASTS=1 PYTHONPATH=. python extra/optimization/get_action_space.py
-    - name: Repo line count < 14600 lines
-      run: MAX_LINE_COUNT=14600 python sz.py
+    - name: Repo line count < 15000 lines
+      run: MAX_LINE_COUNT=15000 python sz.py
 
   fuzzing:
     name: Fuzzing

test/external/external_llm_eval.py

@@ -0,0 +1,35 @@
+# eval for tinygrad.apps.llm
+import pyarrow.parquet as pq
+from tinygrad.helpers import fetch, colored
+from tinygrad.apps.llm import Transformer, SimpleTokenizer, models
+from tinygrad import Tensor
+
+if __name__ == "__main__":
+  dat = fetch("https://huggingface.co/datasets/allenai/ai2_arc/resolve/main/ARC-Challenge/test-00000-of-00001.parquet")
+  table = pq.read_table(dat)
+
+  model, kv = Transformer.from_gguf(Tensor.from_url(models["1B"]), max_context=4096)
+
+  tok = SimpleTokenizer(kv["tokenizer.ggml.tokens"])
+  bos_id: int = kv['tokenizer.ggml.bos_token_id']
+  eos_id: int = kv['tokenizer.ggml.eos_token_id']
+
+  num_correct, num_answered = 0, 0
+  total_questions = len(table["question"])
+  for question, choices, answer in zip(table["question"], table["choices"], table["answerKey"]):
+    phrasing = f"Question: {question}\n\n" + \
+               '\n'.join([f"{k}) {v}" for k,v in zip(choices['label'], choices['text'])]) +\
+               "\n\nReply with the letter of the correct answer only."
+    try:
+      ids = [bos_id] + tok.role("user") + tok.encode(phrasing) + [eos_id] + tok.role("assistant") + tok.encode("Answer: ")
+    except RuntimeError:
+      # TODO: fix the tokenizer
+      pass
+    next_id = next(model.generate(ids))
+    correct, given = answer.as_py().strip(), tok.decode([next_id]).strip()
+    num_correct += correct == given
+    num_answered += 1
+    print(f"{num_answered:4d}/{total_questions:4d}  "+\
+          f"Correct Answer: {correct}  "+\
+          f"Given Answer: {colored(given, 'green' if correct==given else 'red')}  "+\
+          f"Percent: {num_correct*100.0/num_answered:.2f}%")

tinygrad/apps/llm.py

@@ -0,0 +1,179 @@
+from __future__ import annotations
+import sys, argparse
+from tinygrad import Tensor, nn, UOp, TinyJit, getenv
+
+class SimpleTokenizer:
+  def __init__(self, vocab: list[str]):
+    self.vocab: list[str] = vocab
+    self.biggest_token: int = max(map(len, vocab))
+    self.token_to_id: dict[str, int] = {tok: i for i, tok in enumerate(vocab)}
+    self.replace_space = "Ġ"
+    self.replace_newline = "Ċ"
+
+  def encode(self, text:str) -> list[int]:
+    s = text.replace(" ", self.replace_space).replace("\n", self.replace_newline)
+    out: list[int] = []
+    i = 0
+    while i < len(s):
+      j = min(i+self.biggest_token, len(s))
+      while i < j and (tid:=self.token_to_id.get(s[i:j])) is None: j -= 1
+      if tid is None: raise RuntimeError(f"token not found in {s}")
+      assert tid is not None, f"token not found in {s}"
+      out.append(tid)
+      i = j
+    return out
+
+  def decode(self, ids: list[int]) -> str:
+    return ''.join(self.vocab[tid] for tid in ids).replace(self.replace_space, " ").replace(self.replace_newline, "\n")
+
+  def role(self, role:str):
+    return [t for x in ["<|start_header_id|>", role, "<|end_header_id|>\n\n"] for t in self.encode(x)]  # llama style
+
+def apply_rope(x:Tensor, start_pos:int|UOp, base:int=10000):
+  B, H, T, Hd = x.shape
+  # NOTE: this is usually in a RoPE cache, but tinygrad JIT should prune it outside the kernel
+  # TODO: make it do that
+  freq = base ** (-Tensor.arange(0, 1, 2/Hd, dtype='float32'))
+  angles = Tensor.arange(start_pos, start_pos+T, dtype='float32')[None, None, :, None] * freq
+  cos, sin = angles.cos(), angles.sin()
+  x = x.reshape(B, H, T, Hd // 2, 2)    # split into pairs
+  y1 = x[..., 0] * cos - x[..., 1] * sin
+  y2 = x[..., 0] * sin + x[..., 1] * cos
+  return Tensor.stack(y1, y2, dim=-1).reshape(B, H, T, Hd)
+
+class TransformerBlock:
+  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int, norm_eps:float, max_context:int=0):
+    self.n_heads      = n_heads
+    self.n_kv_heads   = n_kv_heads
+    self.head_dim     = dim // n_heads
+    self.max_context  = max_context
+
+    # --- attention projections (all linear, bias-free) ------------------
+    kv_proj_out      = self.head_dim * n_kv_heads    # Llama-3 uses the same dim for K/V
+    self.attn_q      = nn.Linear(dim, dim,         bias=False)
+    self.attn_k      = nn.Linear(dim, kv_proj_out, bias=False)
+    self.attn_v      = nn.Linear(dim, kv_proj_out, bias=False)
+    self.attn_output = nn.Linear(dim, dim,         bias=False)
+
+    # --- RMSNorms --------------------------------------------------------
+    self.attn_norm   = nn.RMSNorm(dim, norm_eps)
+    self.ffn_norm    = nn.RMSNorm(dim, norm_eps)
+
+    # --- feed-forward ----------------------------------------------------
+    self.ffn_gate    = nn.Linear(dim, hidden_dim, bias=False)
+    self.ffn_up      = nn.Linear(dim, hidden_dim, bias=False)
+    self.ffn_down    = nn.Linear(hidden_dim, dim, bias=False)
+
+  def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
+    x_norm = self.attn_norm(x)                       # (B,T,D)
+    q, k, v = self.attn_q(x_norm), self.attn_k(x_norm), self.attn_v(x_norm)
+
+    B, T, _ = x.shape
+    q = q.reshape(B, T, self.n_heads,    self.head_dim).transpose(1, 2)  # (B,H,T,Hd)
+    k = k.reshape(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)  # (B,KvH,T,Hd)
+    v = v.reshape(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)  # (B,KvH,T,Hd)
+
+    q = apply_rope(q, start_pos)
+    k = apply_rope(k, start_pos)
+
+    # TODO: remove these kv cache realizes
+    if not hasattr(self, "cache_kv"):
+      self.cache_kv = Tensor.zeros(2, B, self.n_kv_heads, self.max_context, self.head_dim, dtype=k.dtype, device=k.device).contiguous().realize()
+    self.cache_kv[:, :, :, start_pos:start_pos+T, :].assign(Tensor.stack(k, v)).realize()  # type: ignore
+    k = self.cache_kv[0, :, :, 0:start_pos+T, :]
+    v = self.cache_kv[1, :, :, 0:start_pos+T, :]
+
+    # NOTE: this mask is causal_lower_right, not the causal_upper_left generated by is_casual = True
+    mask = Tensor.full((1, 1, T, start_pos+T), float("-inf"), dtype=x.dtype, device=x.device).triu(start_pos+1) if T > 1 else None
+    attn = q.scaled_dot_product_attention(k, v, attn_mask=mask, enable_gqa=True)     # (B,H,T,Hd)
+    attn = attn.transpose(1, 2).reshape(B, T, -1)                                    # back to (B,T,D)
+    attn = self.attn_output(attn)
+    return x + attn
+
+  def _feed_forward(self, h: Tensor) -> Tensor:
+    h_norm = self.ffn_norm(h)
+    gated  = self.ffn_gate(h_norm).silu() * self.ffn_up(h_norm)
+    return h + self.ffn_down(gated)
+
+  def __call__(self, x: Tensor, start_pos: int|UOp):
+    return self._feed_forward(self._attention(x, start_pos))
+
+class Transformer:
+  def __init__(self, *, num_blocks, dim, hidden_dim, n_heads, n_kv_heads, norm_eps, vocab_size, max_context):
+    self.blk = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, max_context) for _ in range(num_blocks)]
+    self.token_embd  = nn.Embedding(vocab_size, dim)
+    self.output_norm = nn.RMSNorm(dim, norm_eps)
+    self.output = nn.Linear(dim, vocab_size, bias=False)
+    self.max_context = max_context
+    # JIT is used if T=1 and start_pos is a UOp. TODO: make this not needed by including T in the JIT and making start_pos always a UOp
+    self.forward_jit = TinyJit(self.forward)
+
+  def forward(self, tokens:Tensor, start_pos:int|UOp) -> Tensor:
+    x = self.token_embd(tokens)                           # (B, T, D)
+    for block in self.blk: x = block(x, start_pos)
+    # TODO: add temperature
+    return self.output(self.output_norm(x))[:, -1, :].softmax(-1).argmax(-1, keepdim=True)
+
+  def __call__(self, tokens:Tensor, start_pos:int|UOp=0) -> Tensor:
+    return (self.forward_jit if getenv("JIT", 1) and tokens.shape[1] == 1 and isinstance(start_pos, UOp) else self.forward)(tokens, start_pos)
+
+  @staticmethod
+  def from_gguf(gguf:Tensor, max_context:int|None=None) -> tuple[Transformer, dict]:
+    # TODO: remove the need for copy to default device
+    kv, state_dict = nn.state.gguf_load(gguf.to(None))
+
+    # all state items should be float16, not float32
+    state_dict = {k:v.cast('float16') for k,v in state_dict.items()}
+
+    # some models like Llama 3.2 don't have an output.weight, they just tie to the token_embd.weight
+    if 'output.weight' not in state_dict: state_dict['output.weight'] = state_dict['token_embd.weight']
+
+    arch = kv['general.architecture']
+    max_context = min(max_context, kv[f'{arch}.context_length']) if max_context is not None else kv[f'{arch}.context_length']
+    model = Transformer(num_blocks=kv[f'{arch}.block_count'], dim=kv[f'{arch}.embedding_length'], hidden_dim=kv[f'{arch}.feed_forward_length'],
+                        n_heads=kv[f'{arch}.attention.head_count'], n_kv_heads=kv[f'{arch}.attention.head_count_kv'],
+                        norm_eps=kv[f'{arch}.attention.layer_norm_rms_epsilon'], vocab_size=len(kv['tokenizer.ggml.tokens']), max_context=max_context)
+    nn.state.load_state_dict(model, state_dict, verbose=False, consume=True, realize=False)  # NOTE: rope_freqs.weight (32,) is unused
+    return model, kv
+
+  def generate(self, tokens:list[int], start_pos=0):
+    v_start_pos = UOp.variable("start_pos", 1, self.max_context-1)
+    start_pos = 0
+    t = Tensor([tokens[start_pos:]], dtype="int32")
+    self.forward_jit.reset()  # TODO: why is this required? root cause the issue and make it not be needed
+    while len(tokens) < self.max_context:
+      t = self(t, v_start_pos.bind(start_pos) if getenv("SYM", 1) and start_pos != 0 and t.shape[-1] == 1 else start_pos)
+      next_id = int(t.item())
+      tokens.append(next_id)
+      start_pos = len(tokens) - 1
+      yield next_id
+
+models = {
+  "1B": "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q6_K.gguf",
+  "3B": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-Q6_K.gguf",
+  "3B_f16": "https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/Llama-3.2-3B-Instruct-f16.gguf",
+  "8B": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q8_0.gguf",
+}
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--size", choices=list(models.keys()), default=list(models.keys())[0], help="Model size")
+  parser.add_argument("--max_context", type=int, default=4096, help="Max Context Length")
+  args = parser.parse_args()
+
+  # load the model
+  model, kv = Transformer.from_gguf(Tensor.from_url(models[args.size]), args.max_context)
+
+  # extract some metadata
+  tok = SimpleTokenizer(kv["tokenizer.ggml.tokens"])
+  bos_id: int = kv['tokenizer.ggml.bos_token_id']
+  eos_id: int = kv['tokenizer.ggml.eos_token_id']
+
+  ids: list[int] = [bos_id]
+  while 1:
+    start_pos = len(ids) - 1
+    ids += tok.role("user") + tok.encode(input('>>> ')) + [eos_id] + tok.role("assistant")
+    for next_id in model.generate(ids, start_pos):
+      sys.stdout.write(tok.decode([next_id]) if next_id != eos_id else "\n\n")
+      sys.stdout.flush()
+      if next_id == eos_id: break