CANN: Uses yarn_ramp cache in ROPE (#17725)

* webui: apply webui_settings on first load

The webui_settings from /props were not applied on initial load
when default_generation_settings.params was null

Now syncs whenever serverProps is available, regardless of params,
works for both single-model and router modes

* chore: update webui build output

common/arg.cpp

@@ -2087,7 +2087,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "override tensor buffer type", [](common_params & params, const std::string & value) {
             parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
         }
-    ));
+    ).set_env("LLAMA_ARG_OVERRIDE_TENSOR"));
     add_opt(common_arg(
         {"-otd", "--override-tensor-draft"}, "<tensor name pattern>=<buffer type>,...",
         "override tensor buffer type for draft model", [](common_params & params, const std::string & value) {

examples/model-conversion/Makefile

@@ -25,6 +25,8 @@ define quantize_model
 	@echo "Export the quantized model path to $(2) variable in your environment"
 endef
 
+DEVICE ?= auto
+
 ###
 ### Casual Model targets/recipes
 ###
@@ -53,7 +55,7 @@ causal-convert-mm-model:
 
 causal-run-original-model:
 	$(call validate_model_path,causal-run-original-model)
-	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py --device "$(DEVICE)"
 
 causal-run-converted-model:
 	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh

examples/model-conversion/scripts/causal/run-org-model.py

@@ -4,149 +4,179 @@ import argparse
 import os
 import sys
 import importlib
+import torch
+import numpy as np
+
 from pathlib import Path
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForImageTextToText, AutoConfig
 
 # Add parent directory to path for imports
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
-
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForImageTextToText, AutoConfig
-import torch
-import numpy as np
 from utils.common import debug_hook
 
-parser = argparse.ArgumentParser(description="Process model with specified path")
-parser.add_argument("--model-path", "-m", help="Path to the model")
-parser.add_argument("--prompt-file", "-f", help="Optional prompt file", required=False)
-parser.add_argument("--verbose", "-v", action="store_true", help="Enable verbose debug output")
-args = parser.parse_args()
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Process model with specified path")
+    parser.add_argument("--model-path", "-m", help="Path to the model")
+    parser.add_argument("--prompt-file", "-f", help="Optional prompt file", required=False)
+    parser.add_argument("--verbose", "-v", action="store_true", help="Enable verbose debug output")
+    parser.add_argument("--device", "-d", help="Device to use (cpu, cuda, mps, auto)", default="auto")
+    return parser.parse_args()
 
-model_path = os.environ.get("MODEL_PATH", args.model_path)
-if model_path is None:
-    parser.error(
-        "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
-    )
+def load_model_and_tokenizer(model_path, device="auto"):
+    print("Loading model and tokenizer using AutoTokenizer:", model_path)
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    multimodal = False
+    full_config = config
 
-### If you want to dump RoPE activations, uncomment the following lines:
-### === START ROPE DEBUG ===
-# from utils.common import setup_rope_debug
-# setup_rope_debug("transformers.models.apertus.modeling_apertus")
-### == END ROPE DEBUG ===
-
-
-print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
-multimodal = False
-full_config = config
-
-print("Model type:       ", config.model_type)
-if "vocab_size" not in config and "text_config" in config:
-    config = config.text_config
-    multimodal = True
-print("Vocab size:       ", config.vocab_size)
-print("Hidden size:      ", config.hidden_size)
-print("Number of layers: ", config.num_hidden_layers)
-print("BOS token id:     ", config.bos_token_id)
-print("EOS token id:     ", config.eos_token_id)
-
-unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
-if unreleased_model_name:
-    model_name_lower = unreleased_model_name.lower()
-    unreleased_module_path = (
-        f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
-    )
-    class_name = f"{unreleased_model_name}ForCausalLM"
-    print(f"Importing unreleased model module: {unreleased_module_path}")
-
-    try:
-        model_class = getattr(
-            importlib.import_module(unreleased_module_path), class_name
-        )
-        model = model_class.from_pretrained(
-            model_path
-        )  # Note: from_pretrained, not fromPretrained
-    except (ImportError, AttributeError) as e:
-        print(f"Failed to import or load model: {e}")
-        exit(1)
-else:
-    if multimodal:
-        model = AutoModelForImageTextToText.from_pretrained(
-            model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=full_config
-        )
+    # Determine device_map based on device argument
+    if device == "cpu":
+        device_map = {"": "cpu"}
+        print("Forcing CPU usage")
+    elif device == "auto":
+        device_map = "auto"
     else:
-        model = AutoModelForCausalLM.from_pretrained(
-            model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=config
+        device_map = {"": device}
+
+    print("Model type:       ", config.model_type)
+    if "vocab_size" not in config and "text_config" in config:
+        config = config.text_config
+        multimodal = True
+
+    print("Vocab size:       ", config.vocab_size)
+    print("Hidden size:      ", config.hidden_size)
+    print("Number of layers: ", config.num_hidden_layers)
+    print("BOS token id:     ", config.bos_token_id)
+    print("EOS token id:     ", config.eos_token_id)
+
+    unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = (
+            f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
         )
+        class_name = f"{unreleased_model_name}ForCausalLM"
+        print(f"Importing unreleased model module: {unreleased_module_path}")
 
-if args.verbose:
-    for name, module in model.named_modules():
-        if len(list(module.children())) == 0:  # only leaf modules
-            module.register_forward_hook(debug_hook(name))
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=config
+            )
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if multimodal:
+            model = AutoModelForImageTextToText.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=full_config
+            )
+        else:
+            model = AutoModelForCausalLM.from_pretrained(
+                    model_path,
+                    device_map=device_map,
+                    offload_folder="offload",
+                    trust_remote_code=True,
+                    config=config
+            )
 
-model_name = os.path.basename(model_path)
-# Printing the Model class to allow for easier debugging. This can be useful
-# when working with models that have not been publicly released yet and this
-# migth require that the concrete class is imported and used directly instead
-# of using AutoModelForCausalLM.
-print(f"Model class: {model.__class__.__name__}")
+    print(f"Model class: {model.__class__.__name__}")
 
-device = next(model.parameters()).device
-if args.prompt_file:
-    with open(args.prompt_file, encoding='utf-8') as f:
-        prompt = f.read()
-elif os.getenv("MODEL_TESTING_PROMPT"):
-    prompt = os.getenv("MODEL_TESTING_PROMPT")
-else:
-    prompt = "Hello, my name is"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+    return model, tokenizer, config
 
-print(f"Input tokens: {input_ids}")
-print(f"Input text: {repr(prompt)}")
-print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+def enable_torch_debugging(model):
+        for name, module in model.named_modules():
+            if len(list(module.children())) == 0:  # only leaf modules
+                module.register_forward_hook(debug_hook(name))
 
-batch_size = 512
+def get_prompt(args):
+    if args.prompt_file:
+        with open(args.prompt_file, encoding='utf-8') as f:
+            return f.read()
+    elif os.getenv("MODEL_TESTING_PROMPT"):
+        return os.getenv("MODEL_TESTING_PROMPT")
+    else:
+        return "Hello, my name is"
 
-with torch.no_grad():
-    past = None
-    outputs = None
-    for i in range(0, input_ids.size(1), batch_size):
-        print(f"Processing chunk with tokens {i} to {i + batch_size}")
-        chunk = input_ids[:, i:i + batch_size]
-        outputs = model(chunk.to(model.device), past_key_values=past, use_cache=True)
-        past = outputs.past_key_values
+def main():
+    args = parse_arguments()
+    model_path = os.environ.get("MODEL_PATH", args.model_path)
+    if model_path is None:
+        print("Error: Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+        sys.exit(1)
 
-    logits = outputs.logits # type: ignore
 
-    # Extract logits for the last token (next token prediction)
-    last_logits = logits[0, -1, :].float().cpu().numpy()
+    model, tokenizer, config = load_model_and_tokenizer(model_path, args.device)
 
-    print(f"Logits shape: {logits.shape}")
-    print(f"Last token logits shape: {last_logits.shape}")
-    print(f"Vocab size: {len(last_logits)}")
+    if args.verbose:
+        enable_torch_debugging(model)
 
-    data_dir = Path("data")
-    data_dir.mkdir(exist_ok=True)
-    bin_filename = data_dir / f"pytorch-{model_name}.bin"
-    txt_filename = data_dir / f"pytorch-{model_name}.txt"
+    model_name = os.path.basename(model_path)
 
-    # Save to file for comparison
-    last_logits.astype(np.float32).tofile(bin_filename)
+    # Iterate over the model parameters (the tensors) and get the first one
+    # and use it to get the device the model is on.
+    device = next(model.parameters()).device
+    prompt = get_prompt(args)
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
 
-    # Also save as text file for easy inspection
-    with open(txt_filename, "w") as f:
-        for i, logit in enumerate(last_logits):
-            f.write(f"{i}: {logit:.6f}\n")
+    print(f"Input tokens: {input_ids}")
+    print(f"Input text: {repr(prompt)}")
+    print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
 
-    # Print some sample logits for quick verification
-    print(f"First 10 logits: {last_logits[:10]}")
-    print(f"Last 10 logits: {last_logits[-10:]}")
+    batch_size = 512
 
-    # Show top 5 predicted tokens
-    top_indices = np.argsort(last_logits)[-5:][::-1]
-    print("Top 5 predictions:")
-    for idx in top_indices:
-        token = tokenizer.decode([idx])
-        print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+    with torch.no_grad():
+        past = None
+        outputs = None
+        for i in range(0, input_ids.size(1), batch_size):
+            print(f"Processing chunk with tokens {i} to {i + batch_size}")
+            chunk = input_ids[:, i:i + batch_size]
+            outputs = model(chunk.to(model.device), past_key_values=past, use_cache=True)
+            past = outputs.past_key_values
 
-    print(f"Saved bin logits to: {bin_filename}")
-    print(f"Saved txt logist to: {txt_filename}")
+        logits = outputs.logits # type: ignore
+
+        # Extract logits for the last token (next token prediction)
+        last_logits = logits[0, -1, :].float().cpu().numpy()
+
+        print(f"Logits shape: {logits.shape}")
+        print(f"Last token logits shape: {last_logits.shape}")
+        print(f"Vocab size: {len(last_logits)}")
+
+        data_dir = Path("data")
+        data_dir.mkdir(exist_ok=True)
+        bin_filename = data_dir / f"pytorch-{model_name}.bin"
+        txt_filename = data_dir / f"pytorch-{model_name}.txt"
+
+        # Save to file for comparison
+        last_logits.astype(np.float32).tofile(bin_filename)
+
+        # Also save as text file for easy inspection
+        with open(txt_filename, "w") as f:
+            for i, logit in enumerate(last_logits):
+                f.write(f"{i}: {logit:.6f}\n")
+
+        # Print some sample logits for quick verification
+        print(f"First 10 logits: {last_logits[:10]}")
+        print(f"Last 10 logits: {last_logits[-10:]}")
+
+        # Show top 5 predicted tokens
+        top_indices = np.argsort(last_logits)[-5:][::-1]
+        print("Top 5 predictions:")
+        for idx in top_indices:
+            token = tokenizer.decode([idx])
+            print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+
+        print(f"Saved bin logits to: {bin_filename}")
+        print(f"Saved txt logist to: {txt_filename}")
+
+if __name__ == "__main__":
+    main()

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2338,19 +2338,19 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
     // Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
     // TODO: acl_yarn_ramp_tensor use rope cache.
     bool                 yarn_ramp_tensor_updated = false;
-    ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
     acl_tensor_ptr       acl_yarn_ramp_tensor;
     if (ext_factor != 0 && (theta_scale_updated || ctx.rope_cache.theta_scale_length != theta_scale_length ||
                             ctx.rope_cache.freq_scale != freq_scale)) {
         yarn_ramp_tensor_updated = true;
-
+        if (ctx.rope_cache.yarn_ramp_cache != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_cache.yarn_ramp_cache));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cache.yarn_ramp_cache, theta_scale_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
         // -rope_yarn_ramp
         // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
         // return MIN(1, MAX(0, y)) - 1;
-        yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
-        void * yarn_ramp_buffer = yarn_ramp_allocator.get();
         acl_yarn_ramp_tensor =
-            ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
+            ggml_cann_create_tensor(ctx.rope_cache.yarn_ramp_cache, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
         float          zero_value = 0, one_value = 1;
         float          denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
         acl_scalar_ptr low              = ggml_cann_create_scalar(&corr_dims[0], aclDataType::ACL_FLOAT);
@@ -2380,8 +2380,10 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
         acl_scalar_ptr freq_scale_1_sc = ggml_cann_create_scalar(&freq_scale_1, aclDataType::ACL_FLOAT);
         GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor.get(), freq_scale_1_sc.get());
         GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor.get(), freq_scale_sc.get(), one.get());
+    } else {
+        acl_yarn_ramp_tensor =
+            ggml_cann_create_tensor(ctx.rope_cache.yarn_ramp_cache, ACL_FLOAT, sizeof(float), theta_scale_ne, theta_scale_nb, 1);
     }
-
     // Step 1.3: update theta_scale_tensor according to ext_factor or freq_scale.
     if (ext_factor != 0) {
         if (theta_scale_updated || yarn_ramp_tensor_updated) {

ggml/src/ggml-cann/common.h

@@ -318,6 +318,9 @@ struct ggml_cann_rope_cache {
         if (position_select_index_host) {
             free(position_select_index_host);
         }
+        if (yarn_ramp_cache) {
+            ACL_CHECK(aclrtFree(yarn_ramp_cache));
+        }
     }
 
     bool equal(int64_t theta_scale_length,
@@ -370,6 +373,7 @@ struct ggml_cann_rope_cache {
     float * theta_scale_exp_host       = nullptr;
     int *   position_select_index_host = nullptr;
     void *  position_select_index      = nullptr;
+    void *  yarn_ramp_cache            = nullptr;
     // sin/cos cache, used only to accelerate first layer on each device
     void *  sin_cache                  = nullptr;
     void *  cos_cache                  = nullptr;

tools/server/server-context.cpp

@@ -2313,6 +2313,12 @@ private:
                         slot.n_prompt_tokens_processed = 0;
 
                         slot.prompt.tokens.keep_first(n_past);
+
+                        // send initial 0% progress update if needed
+                        // this is to signal the client that the request has started processing
+                        if (slot.task->params.stream && slot.task->params.return_progress) {
+                            send_partial_response(slot, {}, true);
+                        }
                     }
 
                     if (!slot.can_split()) {
@@ -2784,6 +2790,12 @@ server_response_reader server_context::get_response_reader() {
 
 server_context_meta server_context::get_meta() const {
     auto tool_use_src = common_chat_templates_source(impl->chat_templates.get(), "tool_use");
+
+    auto bos_id = llama_vocab_bos(impl->vocab);
+    auto eos_id = llama_vocab_eos(impl->vocab);
+    auto bos_token_str = bos_id != LLAMA_TOKEN_NULL ? common_token_to_piece(impl->ctx, bos_id, true) : "";
+    auto eos_token_str = eos_id != LLAMA_TOKEN_NULL ? common_token_to_piece(impl->ctx, eos_id, true) : "";
+
     return server_context_meta {
         /* build_info             */ build_info,
         /* model_name             */ impl->model_name,
@@ -2798,8 +2810,8 @@ server_context_meta server_context::get_meta() const {
         /* chat_template          */ common_chat_templates_source(impl->chat_templates.get()),
         /* chat_template_tool_use */ tool_use_src ? tool_use_src : "",
 
-        /* bos_token_str          */ common_token_to_piece(impl->ctx, llama_vocab_bos(impl->vocab), true),
-        /* eos_token_str          */ common_token_to_piece(impl->ctx, llama_vocab_eos(impl->vocab), true),
+        /* bos_token_str          */ bos_token_str,
+        /* eos_token_str          */ eos_token_str,
         /* fim_pre_token          */ llama_vocab_fim_pre(impl->vocab),
         /* fim_sub_token          */ llama_vocab_fim_suf(impl->vocab),
         /* fim_mid_token          */ llama_vocab_fim_mid(impl->vocab),

tools/server/tests/unit/test_chat_completion.py

@@ -434,8 +434,8 @@ def test_context_size_exceeded_stream():
 @pytest.mark.parametrize(
     "n_batch,batch_count,reuse_cache",
     [
-        (64, 3, False),
-        (64, 1, True),
+        (64, 4, False),
+        (64, 2, True),
     ]
 )
 def test_return_progress(n_batch, batch_count, reuse_cache):
@@ -462,10 +462,18 @@ def test_return_progress(n_batch, batch_count, reuse_cache):
     res = make_cmpl_request()
     last_progress = None
     total_batch_count = 0
+
     for data in res:
         cur_progress = data.get("prompt_progress", None)
         if cur_progress is None:
             continue
+        if total_batch_count == 0:
+            # first progress report must have n_cache == n_processed
+            assert cur_progress["total"] > 0
+            assert cur_progress["cache"] == cur_progress["processed"]
+            if reuse_cache:
+                # when reusing cache, we expect some cached tokens
+                assert cur_progress["cache"] > 0
         if last_progress is not None:
             assert cur_progress["total"] == last_progress["total"]
             assert cur_progress["cache"] == last_progress["cache"]
@@ -473,6 +481,7 @@ def test_return_progress(n_batch, batch_count, reuse_cache):
         total_batch_count += 1
         last_progress = cur_progress
 
+    # last progress should indicate completion (all tokens processed)
     assert last_progress is not None
     assert last_progress["total"] > 0
     assert last_progress["processed"] == last_progress["total"]

tools/server/webui/src/lib/stores/settings.svelte.ts

@@ -294,15 +294,14 @@ class SettingsStore {
 	 * This sets up the default values from /props endpoint
 	 */
 	syncWithServerDefaults(): void {
-		const serverParams = serverStore.defaultParams;
-		if (!serverParams) {
-			console.warn('No server parameters available for initialization');
+		const propsDefaults = this.getServerDefaults();
+
+		if (Object.keys(propsDefaults).length === 0) {
+			console.warn('No server defaults available for initialization');
 
 			return;
 		}
 
-		const propsDefaults = this.getServerDefaults();
-
 		for (const [key, propsValue] of Object.entries(propsDefaults)) {
 			const currentValue = getConfigValue(this.config, key);
 

tools/server/webui/src/routes/+layout.svelte

@@ -119,7 +119,7 @@
 	$effect(() => {
 		const serverProps = serverStore.props;
 
-		if (serverProps?.default_generation_settings?.params) {
+		if (serverProps) {
 			settingsStore.syncWithServerDefaults();
 		}
 	});