Upload model files

README.md

@@ -1,19 +1,29 @@
 ---
 license: mit
 library_name: gguf
+pipeline_tag: time-series-forecasting
+language:
+  - en
+base_model: google/timesfm-2.5-200m-pytorch
+base_model_relation: quantized
+quantized_by: amaye15
 tags:
   - gguf
   - time-series
   - forecasting
+  - zero-shot
+  - probabilistic
+  - transformer
+  - patch-based
   - rust
-base_model: google/timesfm-2.5-200m-pytorch
+inference: false
 ---
 
 # timesfm-rs
 
 Pure Rust converter and inference engine for [google/timesfm-2.5-200m-pytorch](https://huggingface.co/google/timesfm-2.5-200m-pytorch).
 
-Produces GGUF v3 files and runs native forecasting — no Python required.
+Pre-converted GGUF files are available at [amaye15/timesfm-gguf](https://huggingface.co/amaye15/timesfm-gguf). Produces GGUF v3 files and runs native forecasting — no Python required.
 
 ## Build
 
@@ -78,9 +88,9 @@ Output is JSON in an OpenAI-compatible forecast format with point forecast and a
     "forecast": {
       "point": [2.1, 2.3, 2.5, "..."],
       "quantiles": {
-        "0.1": [1.8, 2.0, 2.2, "..."],
-        "0.5": [2.1, 2.3, 2.5, "..."],
-        "0.9": [2.4, 2.6, 2.8, "..."]
+        "0.10": [1.8, 2.0, 2.2, "..."],
+        "0.50": [2.1, 2.3, 2.5, "..."],
+        "0.90": [2.4, 2.6, 2.8, "..."]
       }
     },
     "finish_reason": "stop"

timesfm-rs/.cargo/config.toml

@@ -0,0 +1,2 @@
+[build]
+rustflags = ["-C", "target-cpu=native"]

timesfm-rs/Cargo.toml

@@ -26,7 +26,11 @@ candle-core  = { version = "0.8", features = ["accelerate"] }
 candle-nn    = { version = "0.8", features = ["accelerate"] }
 
 [profile.release]
-opt-level = 3
+opt-level     = 3
+lto           = "fat"
+codegen-units = 1
+strip         = true
+panic         = "abort"
 
 
 [features]

timesfm-rs/scripts/compare_python.py

@@ -18,8 +18,8 @@ TIMESFM_SRC = os.path.join(REPO_ROOT, "..", "timesfm", "src")
 if os.path.isdir(TIMESFM_SRC):
     sys.path.insert(0, TIMESFM_SRC)
 
-QUANTILE_LABELS = ["point", "q0.1", "q0.2", "q0.3", "q0.4", "q0.5", "q0.6", "q0.7", "q0.8", "q0.9"]
-QUANTILE_KEYS   = ["0.1", "0.2", "0.3", "0.4", "0.5", "0.6", "0.7", "0.8", "0.9"]
+QUANTILE_LABELS = ["point", "q0.10", "q0.20", "q0.30", "q0.40", "q0.50", "q0.60", "q0.70", "q0.80", "q0.90"]
+QUANTILE_KEYS   = ["0.10", "0.20", "0.30", "0.40", "0.50", "0.60", "0.70", "0.80", "0.90"]
 
 
 def make_context(n=128):

timesfm-rs/src/infer/mod.rs

@@ -201,7 +201,7 @@ impl TimesFMModel {
             &content, &mut reader, "out_point", D_MODEL, D_MODEL, false, &device,
         )?;
 
-        let rope = RopeCache::new(HEAD_DIM, MAX_SEQ, ROPE_THETA);
+        let rope = RopeCache::new(HEAD_DIM, MAX_SEQ, ROPE_THETA, &device)?;
 
         Ok(Self { device, rope, tokenizer, blocks, out_point, causal_mask_cache: Mutex::new(HashMap::new()) })
     }
@@ -395,12 +395,12 @@ impl TimesFMModel {
         let cached_len = rope_offset;
         let decode_mask = make_decode_mask(M_PATCHES, cached_len, &self.device)?;
         for (li, block) in self.blocks.iter().enumerate() {
+            // decode_block_kv returns the extended K/V (old cache + new M_PATCHES).
+            // Store directly — no second Tensor::cat needed.
             let (h_out, k_new, v_new) =
                 self.decode_block_kv(hidden, block, &kv_cache[li], rope_offset, &decode_mask)?;
             hidden = h_out;
-            let k_full = Tensor::cat(&[&kv_cache[li].0, &k_new], 2)?;
-            let v_full = Tensor::cat(&[&kv_cache[li].1, &v_new], 2)?;
-            kv_cache[li] = (k_full, v_full);
+            kv_cache[li] = (k_new, v_new);
         }
         let out_seq = hidden.squeeze(0)?;
         forward_residual_block(&out_seq, &self.out_point, false)
@@ -452,7 +452,9 @@ impl TimesFMModel {
         let k = k.permute([0, 2, 1, 3])?.contiguous()?;
         let v = v.permute([0, 2, 1, 3])?.contiguous()?;
 
-        let k_full = Tensor::cat(&[&cache.0, &k], 2)?;    // [1, N_HEADS, cached+M, HEAD_DIM]
+        // Extend cache: [1, N_HEADS, cached+M, HEAD_DIM].
+        // Return k_full/v_full so the caller can store them directly without a second cat.
+        let k_full = Tensor::cat(&[&cache.0, &k], 2)?;
         let v_full = Tensor::cat(&[&cache.1, &v], 2)?;
 
         let scores = q.matmul(&k_full.transpose(D::Minus1, D::Minus2)?)?;
@@ -460,7 +462,7 @@ impl TimesFMModel {
         let attn_w = candle_nn::ops::softmax(&scores, D::Minus1)?;
         let ctx = attn_w.matmul(&v_full)?;
         let ctx = ctx.permute([0, 2, 1, 3])?.contiguous()?.reshape((1, M_PATCHES, D_MODEL))?;
-        Ok((linear(&ctx, &w.out_w, None)?, k, v))
+        Ok((linear(&ctx, &w.out_w, None)?, k_full, v_full))
     }
 
     fn forward_ffn(&self, x: Tensor, w: &FfnW) -> Result<Tensor> {

timesfm-rs/src/infer/rope.rs

@@ -1,14 +1,13 @@
 use anyhow::Result;
-use candle_core::{DType, Tensor, D};
+use candle_core::{DType, Device, Tensor, D};
 
 pub struct RopeCache {
-    cos: Vec<f32>,
-    sin: Vec<f32>,
-    head_dim: usize,
+    cos_t: Tensor, // [max_seq, head_dim]
+    sin_t: Tensor, // [max_seq, head_dim]
 }
 
 impl RopeCache {
-    pub fn new(head_dim: usize, max_seq: usize, theta: f64) -> Self {
+    pub fn new(head_dim: usize, max_seq: usize, theta: f64, device: &Device) -> Result<Self> {
         let half = head_dim / 2;
         let inv_freq: Vec<f64> = (0..half)
             .map(|i| 1.0 / theta.powf(2.0 * i as f64 / head_dim as f64))
@@ -26,7 +25,9 @@ impl RopeCache {
                 sin[p * head_dim + half + i] = s;
             }
         }
-        Self { cos, sin, head_dim }
+        let cos_t = Tensor::from_vec(cos, (max_seq, head_dim), device)?;
+        let sin_t = Tensor::from_vec(sin, (max_seq, head_dim), device)?;
+        Ok(Self { cos_t, sin_t })
     }
 
     /// Apply RoPE to x with shape [b, seq, n_heads, head_dim].
@@ -38,16 +39,10 @@ impl RopeCache {
         let hd = dims[3];
         let half = hd / 2;
         let dtype = x.dtype();
-        let device = x.device();
 
-        let cos_slice: Vec<f32> = self.cos[start_pos * hd..(start_pos + seq) * hd].to_vec();
-        let sin_slice: Vec<f32> = self.sin[start_pos * hd..(start_pos + seq) * hd].to_vec();
-
-        // [seq, hd] → [1, seq, 1, hd] for broadcasting over [b, seq, nh, hd]
-        let cos_t = Tensor::from_vec(cos_slice, (seq, hd), device)?
-            .unsqueeze(0)?.unsqueeze(2)?;
-        let sin_t = Tensor::from_vec(sin_slice, (seq, hd), device)?
-            .unsqueeze(0)?.unsqueeze(2)?;
+        // Zero-copy narrow: produces a view into the pre-built Tensor
+        let cos_t = self.cos_t.narrow(0, start_pos, seq)?.unsqueeze(0)?.unsqueeze(2)?;
+        let sin_t = self.sin_t.narrow(0, start_pos, seq)?.unsqueeze(0)?.unsqueeze(2)?;
 
         let x32 = x.to_dtype(DType::F32)?;
         let x1 = x32.narrow(D::Minus1, 0, half)?;
@@ -55,7 +50,7 @@ impl RopeCache {
         let cos1 = cos_t.narrow(D::Minus1, 0, half)?;
         let sin1 = sin_t.narrow(D::Minus1, 0, half)?;
 
-        let first = (x1.broadcast_mul(&cos1)? - x2.broadcast_mul(&sin1)?)?;
+        let first  = (x1.broadcast_mul(&cos1)? - x2.broadcast_mul(&sin1)?)?;
         let second = (x2.broadcast_mul(&cos1)? + x1.broadcast_mul(&sin1)?)?;
 
         Ok(Tensor::cat(&[&first, &second], D::Minus1)?.to_dtype(dtype)?)

timesfm-rs/src/main.rs

@@ -155,7 +155,7 @@ fn cmd_infer(gguf_path: &PathBuf) -> Result<()> {
     eprintln!("Loading model from {} …", gguf_path.display());
     let model = TimesFMModel::load(gguf_path)?;
 
-    let quantile_labels = ["0.1", "0.2", "0.3", "0.4", "0.5", "0.6", "0.7", "0.8", "0.9"];
+    let quantile_labels = ["0.10", "0.20", "0.30", "0.40", "0.50", "0.60", "0.70", "0.80", "0.90"];
 
     let mut fc_choices = Vec::new();
     for ctx in &contexts {

timesfm-rs/src/py.rs

@@ -69,7 +69,7 @@ impl TimesFM {
         horizon: usize,
     ) -> PyResult<Py<PyAny>> {
         let contexts = parse_contexts(context)?;
-        let quantile_labels = ["0.1", "0.2", "0.3", "0.4", "0.5", "0.6", "0.7", "0.8", "0.9"];
+        let quantile_labels = ["0.10", "0.20", "0.30", "0.40", "0.50", "0.60", "0.70", "0.80", "0.90"];
         let mut fc_choices = Vec::new();
         for ctx in &contexts {
             if ctx.is_empty() {