Merge branch 'master' into patch-1

Author: karpathy

Makefile

@@ -34,7 +34,7 @@ runomp: run.c
 
 .PHONY: win64
 win64: 
-	x86_64-w64-mingw32-gcc-win32 -Ofast -D_WIN32 -o run.exe -I. run.c win.c
+	x86_64-w64-mingw32-gcc -Ofast -D_WIN32 -o run.exe -I. run.c win.c
 
 # compiles with gnu99 standard flags for amazon linux, coreos, etc. compatibility
 .PHONY: rungnu

README.md

@@ -200,10 +200,14 @@ If your candidate PRs have elements of these it doesn't mean they won't get merg
 - [llama2.go](https://github.com/haormj/llama2.go) by @haormj: a Go port of this project
 - [llama2.go](https://github.com/saracen/llama2.go) by @saracen: a Go port of this project
 - [llama2.c-android](https://github.com/Manuel030/llama2.c-android): by @Manuel030: adds Android binaries of this project
+- [llama2.c-android-wrapper](https://github.com/celikin/llama2.c-android-wrapper): by @celikin: added JNI wrapper, PoC
 - [llama2.cpp](https://github.com/leloykun/llama2.cpp) by @leloykun: a C++ port of this project
 - [llama2.js](https://github.com/epicure/llama2.js) by @epicure: a JavaScript port of this project
 - [llama2.zig](https://github.com/cgbur/llama2.zig) by @cgbur: A Zig port of this project
 - [llama2.jl](https://github.com/juvi21/llama2.jl) by @juvi21: a Julia port of this project
+- [llama2.c - Llama 2 Everywhere](https://github.com/trholding/llama2.c) by @trholding: Standalone, Bootable & Portable Binary Llama 2
+- [llama2.rs](https://github.com/leo-du/llama2.rs) by @leo-du: A Rust port of this project
+- [llama2.scala](https://github.com/jrudolph/llama2.scala) by @jrudolph: a Scala port of this project
 
 ## unsorted todos
 

build_msvc.bat

@@ -1 +1 @@
-cl.exe /Ox /openmp /I. run.c win.c
+cl.exe /fp:fast /Ox /openmp /I. run.c win.c 

export_meta_llama_bin.py

@@ -56,12 +56,12 @@ def serialize(key):
 
     # final rmsnorm
     serialize('norm.weight')
-    # freqs_cis
-    freqs_cis = precompute_freqs_cis(p['dim'] // p['n_heads'], p['max_seq_len'] * 2)
-    state_dict['freqs_cis.real'] = freqs_cis.real[:p['max_seq_len']]
-    state_dict['freqs_cis.imag'] = freqs_cis.imag[:p['max_seq_len']]
-    serialize('freqs_cis.real')
-    serialize('freqs_cis.imag')
+    # freqs_cos, freqs_sin
+    freqs_cos, freqs_sin = precompute_freqs_cis(p['dim'] // p['n_heads'], p['max_seq_len'] * 2)
+    state_dict['freqs_cos'] = freqs_cos[:p['max_seq_len']]
+    state_dict['freqs_sin'] = freqs_sin[:p['max_seq_len']]
+    serialize('freqs_cos')
+    serialize('freqs_sin')
 
     # finally write the output weights
     serialize('output.weight')

model.py

@@ -40,9 +40,9 @@ def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
     freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
     t = torch.arange(end, device=freqs.device)  # type: ignore
     freqs = torch.outer(t, freqs).float()  # type: ignore
-    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
-    return freqs_cis
-
+    freqs_cos = torch.cos(freqs)  # real part
+    freqs_sin = torch.sin(freqs)  # imaginary part
+    return freqs_cos, freqs_sin
 
 def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
     ndim = x.ndim
@@ -51,17 +51,31 @@ def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
     shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
     return freqs_cis.view(*shape)
 
-
 def apply_rotary_emb(
     xq: torch.Tensor,
     xk: torch.Tensor,
-    freqs_cis: torch.Tensor,
+    freqs_cos: torch.Tensor,
+    freqs_sin: torch.Tensor
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
-    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
-    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
-    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
-    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+
+    # reshape xq and xk to match the complex representation
+    xq_r, xq_i = xq.float().reshape(*xq.shape[:-1], -1, 2).unbind(-1)
+    xk_r, xk_i = xk.float().reshape(*xk.shape[:-1], -1, 2).unbind(-1)
+    
+    # reshape freqs_cos and freqs_sin for broadcasting
+    freqs_cos = reshape_for_broadcast(freqs_cos, xq_r)
+    freqs_sin = reshape_for_broadcast(freqs_sin, xq_r)
+
+    # apply rotation using real numbers
+    xq_out_r = xq_r * freqs_cos - xq_i * freqs_sin
+    xq_out_i = xq_r * freqs_sin + xq_i * freqs_cos
+    xk_out_r = xk_r * freqs_cos - xk_i * freqs_sin
+    xk_out_i = xk_r * freqs_sin + xk_i * freqs_cos
+
+    # flatten last two dimensions
+    xq_out = torch.stack([xq_out_r, xq_out_i], dim=-1).flatten(3)
+    xk_out = torch.stack([xk_out_r, xk_out_i], dim=-1).flatten(3)
+
     return xq_out.type_as(xq), xk_out.type_as(xk)
 
 def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
@@ -103,7 +117,8 @@ def __init__(self, args: ModelArgs):
     def forward(
         self,
         x: torch.Tensor,
-        freqs_cis: torch.Tensor,
+        freqs_cos: torch.Tensor,
+        freqs_sin: torch.Tensor,
     ):
         bsz, seqlen, _ = x.shape
 
@@ -114,7 +129,7 @@ def forward(
         xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
 
         # RoPE relative positional embeddings
-        xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+        xq, xk = apply_rotary_emb(xq, xk, freqs_cos, freqs_sin)
 
         # grouped multiquery attention: expand out keys and values
         xk = repeat_kv(xk, self.n_rep)  # (bs, seqlen, n_local_heads, head_dim)
@@ -176,8 +191,8 @@ def __init__(self, layer_id: int, args: ModelArgs):
         self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
         self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
 
-    def forward(self, x, freqs_cis):
-        h = x + self.attention.forward(self.attention_norm(x), freqs_cis)
+    def forward(self, x, freqs_cos, freqs_sin):
+        h = x + self.attention.forward(self.attention_norm(x), freqs_cos, freqs_sin)
         out = h + self.feed_forward.forward(self.ffn_norm(h))
         return out
 
@@ -201,8 +216,9 @@ def __init__(self, params: ModelArgs):
         self.tok_embeddings.weight = self.output.weight # https://paperswithcode.com/method/weight-tying
 
         # some useful precompute for the RoPE relative positional embeddings. TODO why * 2 here? confuse
-        freqs_cis = precompute_freqs_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len * 2)
-        self.register_buffer("freqs_cis", freqs_cis, persistent=False)
+        freqs_cos, freqs_sin = precompute_freqs_cis(self.params.dim // self.params.n_heads, self.params.max_seq_len * 2)
+        self.register_buffer("freqs_cos", freqs_cos, persistent=False)
+        self.register_buffer("freqs_sin", freqs_sin, persistent=False)
 
         # init all weights
         self.apply(self._init_weights)
@@ -223,10 +239,11 @@ def forward(self, tokens, targets=None):
         _bsz, seqlen = tokens.shape
         h = self.tok_embeddings(tokens)
         h = self.dropout(h)
-        freqs_cis = self.freqs_cis[:seqlen]
+        freqs_cos = self.freqs_cos[:seqlen]
+        freqs_sin = self.freqs_sin[:seqlen]
 
         for layer in self.layers:
-            h = layer(h, freqs_cis)
+            h = layer(h, freqs_cos, freqs_sin)
         h = self.norm(h)
 
         if targets is not None:
@@ -359,8 +376,8 @@ def serialize(t):
         serialize(self.norm.weight)
         # note: no need to write final classifier weights due to weight sharing
         # freqs_cis
-        serialize(self.freqs_cis.real[:p.max_seq_len])
-        serialize(self.freqs_cis.imag[:p.max_seq_len])
+        serialize(self.freqs_cos[:p.max_seq_len])
+        serialize(self.freqs_sin[:p.max_seq_len])
 
         # write to binary file
         f.close()