Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+KoHRM-Text-1.4B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+KoHRM-Text-1.4B-Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+KoHRM-Text-1.4B-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
+KoHRM-Text-1.4B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

KoHRM-Text-1.4B-BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5c66f994327c1e2f05b33b0a2ff798a1d05f8b905b7f93943e101bca06c8b0a
+size 2774976832

KoHRM-Text-1.4B-Q5_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90f47f54bd7cf545583a2be43a9d0c971cf6112ff16261e2e926cfabe2e9e35a
+size 1007618400

KoHRM-Text-1.4B-Q6_K.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd54d24344e842c3cd0f261e4b740c42c0ec78ed0b3414cdb8b2ac5022b7fb8a
+size 1142147424

KoHRM-Text-1.4B-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dae86207987804c7e8fc34fcba0d78ae2e54cd8563e907e9e5aea8442f7300c
+size 1477364064

README.md

@@ -0,0 +1,126 @@
+---
+license: apache-2.0
+language:
+- ko
+- en
+library_name: gguf
+pipeline_tag: text-generation
+base_model: LLM-OS-Models/KoHRM-Text-1.4B
+base_model_relation: quantized
+tags:
+- gguf
+- llama.cpp
+- hrm-text
+- kohrm
+- korean
+- text-generation
+---
+
+# KoHRM-Text-1.4B-GGUF
+
+GGUF exports for [LLM-OS-Models/KoHRM-Text-1.4B](https://huggingface.co/LLM-OS-Models/KoHRM-Text-1.4B).
+
+This is a custom `hrm_text` architecture. Standard upstream `llama.cpp`, Ollama, LM Studio, and other GGUF frontends will not load these files until `hrm_text` support lands upstream. Use the included runtime patch:
+
+```text
+runtime/llama.cpp-hrm_text.patch
+```
+
+The patch is based on the HRM-Text GGUF work from [sinimiini/HRM-Text-1B-GGUF](https://huggingface.co/sinimiini/HRM-Text-1B-GGUF), adapted for KoHRM-Text-1.4B. The KoHRM conversion infers the physical H/L stack depth from safetensors, because the public config reports `num_hidden_layers=32` while the exported tensors are arranged as `H: 16` and `L: 16`.
+
+## Files
+
+| file | type | size | sha256 |
+|---|---:|---:|---|
+| `KoHRM-Text-1.4B-BF16.gguf` | BF16 | 2.6G | `d5c66f994327c1e2f05b33b0a2ff798a1d05f8b905b7f93943e101bca06c8b0a` |
+| `KoHRM-Text-1.4B-Q8_0.gguf` | Q8_0 | 1.4G | `8dae86207987804c7e8fc34fcba0d78ae2e54cd8563e907e9e5aea8442f7300c` |
+| `KoHRM-Text-1.4B-Q6_K.gguf` | Q6_K | 1.1G | `dd54d24344e842c3cd0f261e4b740c42c0ec78ed0b3414cdb8b2ac5022b7fb8a` |
+| `KoHRM-Text-1.4B-Q5_K_M.gguf` | Q5_K_M | 961M | `90f47f54bd7cf545583a2be43a9d0c971cf6112ff16261e2e926cfabe2e9e35a` |
+
+## Build Patched llama.cpp
+
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+git checkout 6a257d44633d4a752183ed778b88d2924d0a6b9d
+git apply /path/to/runtime/llama.cpp-hrm_text.patch
+
+cmake -S . -B build-hrm \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DLLAMA_CURL=OFF \
+  -DGGML_NATIVE=OFF
+
+cmake --build build-hrm --target llama-cli llama-quantize llama-completion llama-results -j 8
+```
+
+## CPU Run
+
+Download a quantized GGUF file:
+
+```bash
+huggingface-cli download LLM-OS-Models/KoHRM-Text-1.4B-GGUF \
+  KoHRM-Text-1.4B-Q5_K_M.gguf \
+  --local-dir .
+```
+
+Run on CPU:
+
+```bash
+./build-hrm/bin/llama-cli \
+  -m ./KoHRM-Text-1.4B-Q5_K_M.gguf \
+  -ngl 0 \
+  -t 4 \
+  -c 512 \
+  -n 80 \
+  --single-turn \
+  --simple-io \
+  -p $'질문: 대한민국 민법상 미성년자의 법률행위 취소에 대해 한 문단으로 설명하라.\n답변:'
+```
+
+Tested locally on CPU with the patched `llama.cpp` build:
+
+```text
+build      : b9263-6a257d446
+model      : KoHRM-Text-1.4B-Q5_K_M.gguf
+Prompt     : 16.9 t/s
+Generation : 7.2 t/s
+```
+
+The full smoke-test logs are in `reports/`.
+
+## Prompt Format
+
+The source model card documents this project-side prompt wrapper:
+
+```text
+<|im_start|><|object_ref_start|>PROMPT<|im_end|>
+```
+
+The public checkpoint is a rolling pretraining-stage model, not a final chat/SFT model. Expect unstable instruction-following behavior and possible immediate end-token generation depending on prompt and sampler settings. The GGUF files here validate conversion and patched CPU runtime support, not final assistant quality.
+
+## GGUF Metadata
+
+Key converted metadata:
+
+```text
+general.architecture = hrm_text
+hrm_text.context_length = 4096
+hrm_text.embedding_length = 1536
+hrm_text.block_count = 128
+hrm_text.layers_per_stack = 16
+hrm_text.h_cycles = 2
+hrm_text.l_cycles = 3
+tokenizer.ggml.model = gpt2
+tokenizer.ggml.pre = qwen2
+tokenizer.ggml.bos_token_id = 2
+tokenizer.ggml.eos_token_id = 35
+tokenizer.ggml.padding_token_id = 0
+```
+
+## Notes
+
+- Source model: `LLM-OS-Models/KoHRM-Text-1.4B`
+- Source revision converted: `c413eee318b28e4f970f1be83698b161e60b3adb`
+- `llama.cpp` base commit used for the patch: `6a257d44633d4a752183ed778b88d2924d0a6b9d`
+- BF16 conversion wrote 259 tensors.
+- `llama-completion` can load the model non-interactively, but in local probes it immediately returned an end token for the tested prompts. `llama-cli --single-turn` produced visible CPU token generation and is the command shown above.

reports/conversion_report.md

@@ -0,0 +1,100 @@
+# KoHRM-Text-1.4B GGUF Conversion Report
+
+Date: 2026-06-13 UTC
+
+## Inputs
+
+- Source repo: `LLM-OS-Models/KoHRM-Text-1.4B`
+- Source revision: `c413eee318b28e4f970f1be83698b161e60b3adb`
+- Files used: `config.json`, `model.safetensors`, `tokenizer.json`, `tokenizer_config.json`
+- Reference implementation: `sinimiini/HRM-Text-1B-GGUF`
+- llama.cpp base commit: `6a257d44633d4a752183ed778b88d2924d0a6b9d`
+
+## Conversion Detail
+
+KoHRM config reports `num_hidden_layers=32`, while tensor names show two physical stacks:
+
+```text
+model.H_module.layers.0..15
+model.L_module.layers.0..15
+```
+
+The converter patch therefore infers `layers_per_stack=16` from safetensors instead of treating config `num_hidden_layers` as per-stack depth.
+
+Key metadata in the BF16 GGUF:
+
+```text
+general.architecture = hrm_text
+hrm_text.block_count = 128
+hrm_text.layers_per_stack = 16
+hrm_text.h_cycles = 2
+hrm_text.l_cycles = 3
+tokenizer.ggml.bos_token_id = 2
+tokenizer.ggml.eos_token_id = 35
+tokenizer.ggml.padding_token_id = 0
+```
+
+BF16 export completed with 259 tensors.
+
+## Commands
+
+BF16:
+
+```bash
+python convert_hf_to_gguf.py /home/work/.artifacts/kohrm_gguf/hf/KoHRM-Text-1.4B \
+  --outfile /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-BF16.gguf \
+  --outtype bf16 \
+  --model-name KoHRM-Text-1.4B
+```
+
+Quantization:
+
+```bash
+./build-hrm/bin/llama-quantize KoHRM-Text-1.4B-BF16.gguf KoHRM-Text-1.4B-Q5_K_M.gguf Q5_K_M 8
+./build-hrm/bin/llama-quantize KoHRM-Text-1.4B-BF16.gguf KoHRM-Text-1.4B-Q6_K.gguf Q6_K 4
+./build-hrm/bin/llama-quantize KoHRM-Text-1.4B-BF16.gguf KoHRM-Text-1.4B-Q8_0.gguf Q8_0 4
+```
+
+CPU smoke test:
+
+```bash
+./build-hrm/bin/llama-cli \
+  -m KoHRM-Text-1.4B-Q5_K_M.gguf \
+  -ngl 0 \
+  -t 4 \
+  -c 512 \
+  -n 80 \
+  --seed 42 \
+  --temp 0.7 \
+  --top-p 0.9 \
+  --single-turn \
+  --simple-io \
+  --no-warmup \
+  --display-prompt \
+  -p $'질문: 대한민국 민법상 미성년자의 법률행위 취소에 대해 한 문단으로 설명하라.\n답변:'
+```
+
+## CPU Smoke Result
+
+The patched runtime loaded `KoHRM-Text-1.4B-Q5_K_M.gguf` and generated tokens on CPU.
+
+Key lines:
+
+```text
+build      : b9263-6a257d446
+model      : KoHRM-Text-1.4B-Q5_K_M.gguf
+Prompt: 16.9 t/s
+Generation: 7.2 t/s
+Exiting...
+```
+
+The generated text is not a legal-quality answer; this source checkpoint is documented as a pretraining-stage model, not an instruction-tuned chat model.
+
+## Artifacts
+
+```text
+d5c66f994327c1e2f05b33b0a2ff798a1d05f8b905b7f93943e101bca06c8b0a  KoHRM-Text-1.4B-BF16.gguf
+90f47f54bd7cf545583a2be43a9d0c971cf6112ff16261e2e926cfabe2e9e35a  KoHRM-Text-1.4B-Q5_K_M.gguf
+dd54d24344e842c3cd0f261e4b740c42c0ec78ed0b3414cdb8b2ac5022b7fb8a  KoHRM-Text-1.4B-Q6_K.gguf
+8dae86207987804c7e8fc34fcba0d78ae2e54cd8563e907e9e5aea8442f7300c  KoHRM-Text-1.4B-Q8_0.gguf
+```

reports/cpu_smoke_test_q5_k_m.txt

@@ -0,0 +1,35 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+0.00.021.160 I llama_completion: llama backend init
+0.00.021.165 I llama_completion: load the model and apply lora adapter, if any
+0.00.021.177 I common_init_result: fitting params to device memory ...
+0.00.021.178 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.319.875 I common_params_fit_impl: projected to use 1601 MiB of host memory vs. 2063714 MiB of total host memory
+0.00.481.459 W load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
+0.00.598.773 W llama_context: n_ctx_seq (512) < n_ctx_train (4096) -- the full capacity of the model will not be utilized
+0.00.788.745 I llama_completion: llama threadpool init, n_threads = 4
+0.00.788.761 I 
+0.00.788.815 I system_info: n_threads = 4 (n_threads_batch = 4) / 224 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.788.817 I 
+0.00.788.930 I sampler seed: 42
+0.00.788.938 I sampler params: 
+	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
+	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
+	top_k = 40, top_p = 0.900, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.700
+	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
+0.00.788.945 I sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
+0.00.788.947 I generate: n_ctx = 512, n_batch = 2048, n_predict = 80, n_keep = 0
+0.00.788.947 I 
+질문: 대한민국 민법상 미성년자의 법률행위 취소에 대해 한 문단으로 설명하라.
+답변: [end of text]
+
+
+0.01.871.068 I common_perf_print:    sampling time =       2.56 ms
+0.01.871.071 I common_perf_print:    samplers time =       0.83 ms /    18 tokens
+0.01.871.088 I common_perf_print:        load time =    1079.66 ms
+0.01.871.091 I common_perf_print: prompt eval time =    1079.42 ms /    17 tokens (   63.50 ms per token,    15.75 tokens per second)
+0.01.871.092 I common_perf_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+0.01.871.093 I common_perf_print:       total time =    1082.33 ms /    18 tokens
+0.01.871.101 I common_perf_print: unaccounted time =       0.35 ms /   0.0 %      (total - sampling - prompt eval - eval) / (total)
+0.01.871.103 I common_perf_print:    graphs reused =          0

reports/cpu_smoke_test_q5_k_m_biased.txt

@@ -0,0 +1,36 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+0.00.002.145 W DEPRECATED: argument '-l' specified multiple times, use comma-separated values instead (only last value will be used)
+0.00.002.148 W DEPRECATED: argument '-l' specified multiple times, use comma-separated values instead (only last value will be used)
+0.00.002.381 I llama_completion: llama backend init
+0.00.002.385 I llama_completion: load the model and apply lora adapter, if any
+0.00.002.393 I common_init_result: fitting params to device memory ...
+0.00.002.394 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.298.352 I common_params_fit_impl: projected to use 1601 MiB of host memory vs. 2063714 MiB of total host memory
+0.00.460.515 W load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
+0.00.568.735 W llama_context: n_ctx_seq (512) < n_ctx_train (4096) -- the full capacity of the model will not be utilized
+0.00.745.245 I llama_completion: llama threadpool init, n_threads = 4
+0.00.745.260 I 
+0.00.745.313 I system_info: n_threads = 4 (n_threads_batch = 4) / 224 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.745.315 I 
+0.00.745.406 I sampler seed: 42
+0.00.745.414 I sampler params: 
+	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
+	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
+	top_k = 40, top_p = 0.920, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.650
+	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
+0.00.745.419 I sampler chain: logits -> logit-bias -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
+0.00.745.421 I generate: n_ctx = 512, n_batch = 2048, n_predict = 80, n_keep = 0
+0.00.745.421 I 
+<|im_start|><|object_ref_start|>환율 변동이 개인 투자에 미치는 영향과 대비 전략은 무엇인가요?<|im_end|> [end of text]
+
+
+0.00.745.437 I common_perf_print:    sampling time =       0.00 ms
+0.00.745.438 I common_perf_print:    samplers time =       0.00 ms /    16 tokens
+0.00.745.440 I common_perf_print:        load time =     199.67 ms
+0.00.745.441 I common_perf_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
+0.00.745.442 I common_perf_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+0.00.745.442 I common_perf_print:       total time =       0.19 ms /     2 tokens
+0.00.745.443 I common_perf_print: unaccounted time =       0.19 ms /  98.5 %      (total - sampling - prompt eval - eval) / (total)
+0.00.745.443 I common_perf_print:    graphs reused =          0

reports/cpu_smoke_test_q5_k_m_biased2.txt

@@ -0,0 +1,34 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+0.00.019.298 I llama_completion: llama backend init
+0.00.019.304 I llama_completion: load the model and apply lora adapter, if any
+0.00.019.314 I common_init_result: fitting params to device memory ...
+0.00.019.315 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.319.746 I common_params_fit_impl: projected to use 1601 MiB of host memory vs. 2063714 MiB of total host memory
+0.00.482.433 W load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
+0.00.600.443 W llama_context: n_ctx_seq (512) < n_ctx_train (4096) -- the full capacity of the model will not be utilized
+0.00.793.072 I llama_completion: llama threadpool init, n_threads = 4
+0.00.793.087 I 
+0.00.793.144 I system_info: n_threads = 4 (n_threads_batch = 4) / 224 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.793.146 I 
+0.00.793.243 I sampler seed: 42
+0.00.793.251 I sampler params: 
+	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
+	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
+	top_k = 40, top_p = 0.920, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.650
+	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
+0.00.793.258 I sampler chain: logits -> logit-bias -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
+0.00.793.259 I generate: n_ctx = 512, n_batch = 2048, n_predict = 80, n_keep = 0
+0.00.793.259 I 
+<|im_start|><|object_ref_start|>환율 변동이 개인 투자에 미치는 영향과 대비 전략은 무엇인가요?<|im_end|> [end of text]
+
+
+0.00.793.278 I common_perf_print:    sampling time =       0.00 ms
+0.00.793.278 I common_perf_print:    samplers time =       0.00 ms /    16 tokens
+0.00.793.281 I common_perf_print:        load time =     209.04 ms
+0.00.793.282 I common_perf_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
+0.00.793.284 I common_perf_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+0.00.793.284 I common_perf_print:       total time =       0.21 ms /     2 tokens
+0.00.793.285 I common_perf_print: unaccounted time =       0.20 ms /  98.1 %      (total - sampling - prompt eval - eval) / (total)
+0.00.793.285 I common_perf_print:    graphs reused =          0

reports/cpu_smoke_test_q5_k_m_ignore_eos.txt

@@ -0,0 +1,34 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+0.00.018.587 I llama_completion: llama backend init
+0.00.018.594 I llama_completion: load the model and apply lora adapter, if any
+0.00.018.605 I common_init_result: fitting params to device memory ...
+0.00.018.606 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.319.421 I common_params_fit_impl: projected to use 1601 MiB of host memory vs. 2063714 MiB of total host memory
+0.00.484.114 W load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
+0.00.594.877 W llama_context: n_ctx_seq (512) < n_ctx_train (4096) -- the full capacity of the model will not be utilized
+0.00.780.020 I llama_completion: llama threadpool init, n_threads = 4
+0.00.780.042 I 
+0.00.780.115 I system_info: n_threads = 4 (n_threads_batch = 4) / 224 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.780.118 I 
+0.00.780.300 I sampler seed: 42
+0.00.780.311 I sampler params: 
+	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
+	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
+	top_k = 40, top_p = 0.920, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.650
+	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
+0.00.780.318 I sampler chain: logits -> logit-bias -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
+0.00.780.321 I generate: n_ctx = 512, n_batch = 2048, n_predict = 80, n_keep = 0
+0.00.780.321 I 
+<|im_start|><|object_ref_start|>환율 변동이 개인 투자에 미치는 영향과 대비 전략은 무엇인가요?<|im_end|> [end of text]
+
+
+0.00.780.350 I common_perf_print:    sampling time =       0.00 ms
+0.00.780.350 I common_perf_print:    samplers time =       0.00 ms /    16 tokens
+0.00.780.354 I common_perf_print:        load time =     202.76 ms
+0.00.780.355 I common_perf_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
+0.00.780.357 I common_perf_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+0.00.780.357 I common_perf_print:       total time =       0.33 ms /     2 tokens
+0.00.780.358 I common_perf_print: unaccounted time =       0.33 ms /  99.1 %      (total - sampling - prompt eval - eval) / (total)
+0.00.780.358 I common_perf_print:    graphs reused =          0

reports/cpu_smoke_test_q5_k_m_llama_cli_single_turn.txt

@@ -0,0 +1,36 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+
+Loading model... 
+
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b9263-6a257d446
+model      : KoHRM-Text-1.4B-Q5_K_M.gguf
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+> 질문: 대한민국 민법상 미성년자의 법률행위 취소에 대해 한 문단으로 설명하라.
+답변:
+
+
+The user wants a single-word, concise explanation of 대한민국 민법상 미성년자의 법률행위 취소, which is a common misconception. The key insight is that 미성년자는 법적으로 일정 기간 동안(보통 10년)을 동안(10년)만 유효한, 즉, 10년 동안(10년)을 한 번 이상, 또는 10년 동안(10년)을 10,000,000,000,
+
+[ Prompt: 16.9 t/s | Generation: 7.2 t/s ]
+
+Exiting...

reports/file_sizes.txt

@@ -0,0 +1,4 @@
+-rw-r--r-- 1 work work 2.6G Jun 13 20:54 /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-BF16.gguf
+-rw-r--r-- 1 work work 961M Jun 13 20:57 /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q5_K_M.gguf
+-rw-r--r-- 1 work work 1.1G Jun 13 20:57 /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q6_K.gguf
+-rw-r--r-- 1 work work 1.4G Jun 13 20:57 /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q8_0.gguf

reports/sha256sums.txt

@@ -0,0 +1,4 @@
+d5c66f994327c1e2f05b33b0a2ff798a1d05f8b905b7f93943e101bca06c8b0a  /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-BF16.gguf
+90f47f54bd7cf545583a2be43a9d0c971cf6112ff16261e2e926cfabe2e9e35a  /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q5_K_M.gguf
+dd54d24344e842c3cd0f261e4b740c42c0ec78ed0b3414cdb8b2ac5022b7fb8a  /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q6_K.gguf
+8dae86207987804c7e8fc34fcba0d78ae2e54cd8563e907e9e5aea8442f7300c  /home/work/.artifacts/kohrm_gguf/gguf/KoHRM-Text-1.4B-Q8_0.gguf

reports/tokenization_probe_q5_k_m.txt

@@ -0,0 +1,53 @@
+warning: no usable GPU found, --gpu-layers option will be ignored
+warning: one possible reason is that llama.cpp was compiled without GPU support
+warning: consult docs/build.md for compilation instructions
+0.00.020.667 I llama_completion: llama backend init
+0.00.020.672 I llama_completion: load the model and apply lora adapter, if any
+0.00.020.681 I common_init_result: fitting params to device memory ...
+0.00.020.682 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.319.399 I common_params_fit_impl: projected to use 1601 MiB of host memory vs. 2063714 MiB of total host memory
+0.00.482.839 W load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
+0.00.591.446 W llama_context: n_ctx_seq (512) < n_ctx_train (4096) -- the full capacity of the model will not be utilized
+0.00.768.838 I llama_completion: llama threadpool init, n_threads = 4
+0.00.768.852 I 
+0.00.768.907 I system_info: n_threads = 4 (n_threads_batch = 4) / 224 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.768.909 I 
+0.00.768.999 I llama_completion: prompt: '<|im_start|><|object_ref_start|>환율 변동이 개인 투자에 미치는 영향과 대비 전략은 무엇인가요?<|im_end|>'
+0.00.769.001 I llama_completion: number of tokens in prompt = 16
+0.00.769.002 I      2 -> '<|im_start|>'
+0.00.769.002 I     32 -> '<|object_ref_start|>'
+0.00.769.003 I  89839 -> '환율'
+0.00.769.003 I  22305 -> ' 변동이'
+0.00.769.003 I   2647 -> ' 개인'
+0.00.769.003 I  58182 -> ' 투자에'
+0.00.769.004 I   8642 -> ' 미치는'
+0.00.769.004 I 101589 -> ' 영향과'
+0.00.769.004 I   8785 -> ' 대비'
+0.00.769.004 I  13153 -> ' 전략'
+0.00.769.005 I    406 -> '은'
+0.00.769.007 I  33789 -> ' 무엇'
+0.00.769.007 I   7156 -> '인가'
+0.00.769.007 I    703 -> '요'
+0.00.769.008 I     76 -> '?'
+0.00.769.008 I      3 -> '<|im_end|>'
+0.00.769.008 I 
+0.00.769.010 I sampler seed: 42
+0.00.769.017 I sampler params: 
+	repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
+	dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1
+	top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
+	mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
+0.00.769.022 I sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
+0.00.769.023 I generate: n_ctx = 512, n_batch = 2048, n_predict = 1, n_keep = 0
+0.00.769.024 I 
+<|im_start|><|object_ref_start|>환율 변동이 개인 투자에 미치는 영향과 대비 전략은 무엇인가요?<|im_end|> [end of text]
+
+
+0.00.769.037 I common_perf_print:    sampling time =       0.00 ms
+0.00.769.038 I common_perf_print:    samplers time =       0.00 ms /    16 tokens
+0.00.769.040 I common_perf_print:        load time =     200.87 ms
+0.00.769.042 I common_perf_print: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)
+0.00.769.042 I common_perf_print:        eval time =       0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
+0.00.769.043 I common_perf_print:       total time =       0.20 ms /     2 tokens
+0.00.769.044 I common_perf_print: unaccounted time =       0.20 ms /  98.0 %      (total - sampling - prompt eval - eval) / (total)
+0.00.769.045 I common_perf_print:    graphs reused =          0

runtime/llama.cpp-hrm_text.patch

@@ -0,0 +1,582 @@
+diff --git a/conversion/__init__.py b/conversion/__init__.py
+index 2c38123df..ecf1be2db 100644
+--- a/conversion/__init__.py
++++ b/conversion/__init__.py
+@@ -95,6 +95,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
+     "HunYuanDenseV1ForCausalLM": "hunyuan",
+     "HunYuanMoEV1ForCausalLM": "hunyuan",
+     "HunYuanVLForConditionalGeneration": "hunyuan",
++    "HrmTextForCausalLM": "hrm_text",
+     "IQuestCoderForCausalLM": "llama",
+     "InternLM2ForCausalLM": "internlm",
+     "InternLM3ForCausalLM": "internlm",
+diff --git a/conversion/hrm_text.py b/conversion/hrm_text.py
+new file mode 100644
+index 000000000..f5002a5c7
+--- /dev/null
++++ b/conversion/hrm_text.py
+@@ -0,0 +1,146 @@
++from __future__ import annotations
++
++import re
++import json
++from pathlib import Path
++
++from typing import Iterable, TYPE_CHECKING
++
++import torch
++from safetensors import safe_open
++
++if TYPE_CHECKING:
++    from torch import Tensor
++
++from .base import ModelBase, TextModel, gguf, logger
++
++
++@ModelBase.register("HrmTextForCausalLM")
++class HrmTextModel(TextModel):
++    model_arch = gguf.MODEL_ARCH.HRM_TEXT
++
++    def __init__(self, *args, **kwargs):
++        super().__init__(*args, **kwargs)
++
++        with open(self.dir_model / "config.json", "r", encoding="utf-8") as f:
++            self.raw_hparams = json.load(f)
++
++        self.layers_per_stack = self._infer_layers_per_stack()
++        self.h_cycles = self.raw_hparams["H_cycles"]
++        self.l_cycles = self.raw_hparams["L_cycles"]
++        self.physical_block_count = self.layers_per_stack * 2
++        self.cache_block_count = self.layers_per_stack * self.h_cycles * (self.l_cycles + 1)
++
++        # GGUF tensors store one physical L stack followed by one physical H stack.
++        # The runtime expands these 32 physical layers across 128 KV-cache slots.
++        self.block_count = self.physical_block_count
++        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
++
++    def _infer_layers_per_stack(self) -> int:
++        """Infer HRM stack depth from tensor names.
++
++        Upstream HRM-Text exports use config["num_hidden_layers"] as the stack
++        depth. KoHRM-Text exports currently store the total physical layer count
++        there, while safetensors still contain separate 0..15 H and L stacks.
++        The GGUF runtime needs the per-stack depth, so tensor names are the
++        source of truth.
++        """
++        stack_depths: dict[str, set[int]] = {"H": set(), "L": set()}
++        for path in sorted(Path(self.dir_model).glob("*.safetensors")):
++            with safe_open(path, framework="pt", device="cpu") as f:
++                for name in f.keys():
++                    match = re.match(r"model\.([LH])_module\.layers\.(\d+)\.", name)
++                    if match:
++                        stack_depths[match.group(1)].add(int(match.group(2)))
++        if stack_depths["H"] and stack_depths["L"]:
++            h_depth = max(stack_depths["H"]) + 1
++            l_depth = max(stack_depths["L"]) + 1
++            if h_depth != l_depth:
++                raise ValueError(f"HRM H/L stack depth mismatch: H={h_depth}, L={l_depth}")
++            return h_depth
++        return int(self.raw_hparams["num_hidden_layers"])
++
++    def set_vocab(self):
++        # HRM-Text ships a Qwen2-style tokenizer.json. Keep it as a plain tokenizer;
++        # do not add a chat template for validation GGUFs.
++        self._set_vocab_gpt2()
++
++    def get_vocab_base_pre(self, tokenizer) -> str:
++        del tokenizer
++        return "qwen2"
++
++    def set_gguf_parameters(self):
++        hp = self.raw_hparams
++        head_dim = hp["head_dim"]
++
++        self.gguf_writer.add_context_length(hp["max_position_embeddings"])
++        self.gguf_writer.add_embedding_length(hp["hidden_size"])
++        self.gguf_writer.add_block_count(self.cache_block_count)
++        self.gguf_writer.add_feed_forward_length(hp["intermediate_size"])
++        self.gguf_writer.add_head_count(hp["num_attention_heads"])
++        self.gguf_writer.add_head_count_kv(hp["num_key_value_heads"])
++        self.gguf_writer.add_key_length(head_dim)
++        self.gguf_writer.add_value_length(head_dim)
++        self.gguf_writer.add_rope_dimension_count(head_dim)
++        self.gguf_writer.add_rope_freq_base(hp.get("rope_theta", 10000.0))
++        self.gguf_writer.add_layer_norm_rms_eps(hp["rms_norm_eps"])
++        self.gguf_writer.add_embedding_scale(hp["embedding_scale"])
++
++        arch = self.gguf_writer.arch
++        self.gguf_writer.add_uint32(gguf.Keys.LLM.HRM_LAYERS_PER_STACK.format(arch=arch), self.layers_per_stack)
++        self.gguf_writer.add_uint32(gguf.Keys.LLM.HRM_H_CYCLES.format(arch=arch), self.h_cycles)
++        self.gguf_writer.add_uint32(gguf.Keys.LLM.HRM_L_CYCLES.format(arch=arch), self.l_cycles)
++        self.gguf_writer.add_bool(gguf.Keys.LLM.HRM_PREFIX_LM.format(arch=arch), bool(hp.get("prefix_lm", False)))
++
++    def _format(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
++        return self.format_tensor_name(key, bid=bid, suffix=suffix)
++
++    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
++        if name == "model.embed_tokens.weight":
++            yield self._format(gguf.MODEL_TENSOR.TOKEN_EMBD), data_torch
++            return
++
++        if name == "lm_head.weight":
++            yield self._format(gguf.MODEL_TENSOR.OUTPUT), data_torch
++            return
++
++        if name == "model.z_L_init":
++            yield self._format(gguf.MODEL_TENSOR.HRM_Z_L_INIT, suffix=""), data_torch
++            return
++
++        match = re.fullmatch(r"model\.([LH])_module\.layers\.(\d+)\.(.+)", name)
++        if match is None:
++            raise ValueError(f"Can not map tensor {name!r}")
++
++        stack, layer_s, tensor_name = match.groups()
++        layer_idx = int(layer_s)
++        if layer_idx >= self.layers_per_stack:
++            raise ValueError(f"Layer index {layer_idx} outside HRM stack size {self.layers_per_stack}")
++
++        physical_bid = layer_idx + (self.layers_per_stack if stack == "H" else 0)
++
++        if tensor_name == "attn.gqkv_proj.weight":
++            gate, q, k, v = torch.chunk(data_torch, 4, dim=0)
++            logger.debug("Split %s as gate, q, k, v", name)
++            yield self._format(gguf.MODEL_TENSOR.ATTN_GATE, physical_bid), gate.contiguous()
++            yield self._format(gguf.MODEL_TENSOR.ATTN_Q, physical_bid), q.contiguous()
++            yield self._format(gguf.MODEL_TENSOR.ATTN_K, physical_bid), k.contiguous()
++            yield self._format(gguf.MODEL_TENSOR.ATTN_V, physical_bid), v.contiguous()
++            return
++
++        if tensor_name == "attn.o_proj.weight":
++            yield self._format(gguf.MODEL_TENSOR.ATTN_OUT, physical_bid), data_torch
++            return
++
++        if tensor_name == "mlp.gate_up_proj.weight":
++            gate, up = torch.chunk(data_torch, 2, dim=0)
++            logger.debug("Split %s as gate, up", name)
++            yield self._format(gguf.MODEL_TENSOR.FFN_GATE, physical_bid), gate.contiguous()
++            yield self._format(gguf.MODEL_TENSOR.FFN_UP, physical_bid), up.contiguous()
++            return
++
++        if tensor_name == "mlp.down_proj.weight":
++            yield self._format(gguf.MODEL_TENSOR.FFN_DOWN, physical_bid), data_torch
++            return
++
++        raise ValueError(f"Can not map tensor {name!r}")
+diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
+index 7fdcf03d7..b84cc8827 100644
+--- a/gguf-py/gguf/constants.py
++++ b/gguf-py/gguf/constants.py
+@@ -144,6 +144,10 @@ class Keys:
+         TOKEN_SHIFT_COUNT                 = "{arch}.token_shift_count"
+         INTERLEAVE_MOE_LAYER_STEP         = "{arch}.interleave_moe_layer_step"
+         FULL_ATTENTION_INTERVAL           = "{arch}.full_attention_interval"
++        HRM_LAYERS_PER_STACK              = "{arch}.layers_per_stack"
++        HRM_H_CYCLES                      = "{arch}.h_cycles"
++        HRM_L_CYCLES                      = "{arch}.l_cycles"
++        HRM_PREFIX_LM                     = "{arch}.prefix_lm"
+         ACTIVATION_SPARSITY_SCALE         = "{arch}.activation_sparsity_scale"
+         ALTUP_ACTIVE_IDX                  = "{arch}.altup.active_idx"
+         ALTUP_NUM_INPUTS                  = "{arch}.altup.num_inputs"
+@@ -410,6 +414,7 @@ class MODEL_ARCH(IntEnum):
+     QWEN3            = auto()
+     QWEN3MOE         = auto()
+     QWEN3NEXT        = auto()
++    HRM_TEXT         = auto()
+     QWEN3VL          = auto()
+     QWEN3VLMOE       = auto()
+     QWEN35           = auto()
+@@ -527,6 +532,7 @@ class MODEL_TENSOR(IntEnum):
+     TOKEN_TYPES          = auto()
+     POS_EMBD             = auto()
+     OUTPUT               = auto()
++    HRM_Z_L_INIT         = auto()
+     DENSE_2_OUT          = auto() # embeddinggemma 2_Dense
+     DENSE_3_OUT          = auto() # embeddinggemma 3_Dense
+     OUTPUT_NORM          = auto()
+@@ -925,6 +931,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
+     MODEL_ARCH.QWEN3:            "qwen3",
+     MODEL_ARCH.QWEN3MOE:         "qwen3moe",
+     MODEL_ARCH.QWEN3NEXT:        "qwen3next",
++    MODEL_ARCH.HRM_TEXT:         "hrm_text",
+     MODEL_ARCH.QWEN3VL:          "qwen3vl",
+     MODEL_ARCH.QWEN3VLMOE:       "qwen3vlmoe",
+     MODEL_ARCH.QWEN35:           "qwen35",
+@@ -1042,6 +1049,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
+     MODEL_TENSOR.POS_EMBD:                  "position_embd",
+     MODEL_TENSOR.OUTPUT_NORM:               "output_norm",
+     MODEL_TENSOR.OUTPUT:                    "output",
++    MODEL_TENSOR.HRM_Z_L_INIT:              "hrm.z_l_init",
+     MODEL_TENSOR.DENSE_2_OUT:                "dense_2", # embeddinggemma 2_Dense
+     MODEL_TENSOR.DENSE_3_OUT:                "dense_3", # embeddinggemma 2_Dense
+     MODEL_TENSOR.ROPE_FREQS:                "rope_freqs",
+@@ -2057,6 +2065,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
+         MODEL_TENSOR.SSM_BETA_ALPHA,
+         MODEL_TENSOR.SSM_OUT
+     ],
++    MODEL_ARCH.HRM_TEXT: [
++        MODEL_TENSOR.TOKEN_EMBD,
++        MODEL_TENSOR.OUTPUT,
++        MODEL_TENSOR.HRM_Z_L_INIT,
++        MODEL_TENSOR.ATTN_Q,
++        MODEL_TENSOR.ATTN_K,
++        MODEL_TENSOR.ATTN_V,
++        MODEL_TENSOR.ATTN_GATE,
++        MODEL_TENSOR.ATTN_OUT,
++        MODEL_TENSOR.FFN_GATE,
++        MODEL_TENSOR.FFN_DOWN,
++        MODEL_TENSOR.FFN_UP,
++    ],
+     MODEL_ARCH.QWEN3VL: [
+         MODEL_TENSOR.TOKEN_EMBD,
+         MODEL_TENSOR.OUTPUT_NORM,
+diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
+index c9eead18a..5b8ee3781 100644
+--- a/src/llama-arch.cpp
++++ b/src/llama-arch.cpp
+@@ -37,6 +37,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+     { LLM_ARCH_QWEN3,            "qwen3"            },
+     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
+     { LLM_ARCH_QWEN3NEXT,        "qwen3next"        },
++    { LLM_ARCH_HRM_TEXT,         "hrm_text"         },
+     { LLM_ARCH_QWEN3VL,          "qwen3vl"          },
+     { LLM_ARCH_QWEN3VLMOE,       "qwen3vlmoe"       },
+     { LLM_ARCH_QWEN35,           "qwen35"           },
+@@ -209,6 +210,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+     { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
+     { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
+     { LLM_KV_FULL_ATTENTION_INTERVAL,           "%s.full_attention_interval"           },
++    { LLM_KV_HRM_LAYERS_PER_STACK,              "%s.layers_per_stack"                  },
++    { LLM_KV_HRM_H_CYCLES,                      "%s.h_cycles"                          },
++    { LLM_KV_HRM_L_CYCLES,                      "%s.l_cycles"                          },
++    { LLM_KV_HRM_PREFIX_LM,                     "%s.prefix_lm"                         },
+ 
+     { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
+     { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
+@@ -346,6 +351,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
+     { LLM_TENSOR_OUTPUT_NORM,                            "output_norm" },
+     { LLM_TENSOR_OUTPUT_NORM_LFM2,                       "token_embd_norm" }, // fix for wrong tensor name
+     { LLM_TENSOR_OUTPUT,                                 "output" },
++    { LLM_TENSOR_HRM_Z_L_INIT,                           "hrm.z_l_init" },
+     { LLM_TENSOR_ROPE_FREQS,                             "rope_freqs" },
+     { LLM_TENSOR_ATTN_NORM,                              "blk.%d.attn_norm" },
+     { LLM_TENSOR_ATTN_Q,                                 "blk.%d.attn_q" },
+@@ -565,6 +571,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+     {LLM_TENSOR_POS_EMBD,                   {LLM_TENSOR_LAYER_INPUT,     GGML_OP_GET_ROWS}},
+     {LLM_TENSOR_TOKEN_TYPES,                {LLM_TENSOR_LAYER_INPUT,     GGML_OP_GET_ROWS}},
+     {LLM_TENSOR_TOKEN_EMBD_NORM,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},  // do the norms on the first layer (not the input layer)
++    {LLM_TENSOR_HRM_Z_L_INIT,               {LLM_TENSOR_LAYER_INPUT,     GGML_OP_MUL}},
+     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+diff --git a/src/llama-arch.h b/src/llama-arch.h
+index 89cf16cc3..fa04b684b 100644
+--- a/src/llama-arch.h
++++ b/src/llama-arch.h
+@@ -41,6 +41,7 @@ enum llm_arch {
+     LLM_ARCH_QWEN3,
+     LLM_ARCH_QWEN3MOE,
+     LLM_ARCH_QWEN3NEXT,
++    LLM_ARCH_HRM_TEXT,
+     LLM_ARCH_QWEN3VL,
+     LLM_ARCH_QWEN3VLMOE,
+     LLM_ARCH_QWEN35,
+@@ -213,6 +214,10 @@ enum llm_kv {
+     LLM_KV_TOKEN_SHIFT_COUNT,
+     LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
+     LLM_KV_FULL_ATTENTION_INTERVAL,
++    LLM_KV_HRM_LAYERS_PER_STACK,
++    LLM_KV_HRM_H_CYCLES,
++    LLM_KV_HRM_L_CYCLES,
++    LLM_KV_HRM_PREFIX_LM,
+ 
+     LLM_KV_ATTENTION_HEAD_COUNT,
+     LLM_KV_ATTENTION_HEAD_COUNT_KV,
+@@ -354,6 +359,7 @@ enum llm_tensor {
+     LLM_TENSOR_DENSE_2_OUT,
+     LLM_TENSOR_DENSE_3_OUT,
+     LLM_TENSOR_OUTPUT,
++    LLM_TENSOR_HRM_Z_L_INIT,
+     LLM_TENSOR_OUTPUT_NORM,
+     LLM_TENSOR_OUTPUT_NORM_LFM2, // fix for wrong tensor name
+     LLM_TENSOR_ROPE_FREQS,
+diff --git a/src/llama-context.cpp b/src/llama-context.cpp
+index ad36c0666..fa80f4260 100644
+--- a/src/llama-context.cpp
++++ b/src/llama-context.cpp
+@@ -2208,6 +2208,9 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
+     if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR || model.arch == LLM_ARCH_QWEN35 || model.arch == LLM_ARCH_QWEN35MOE) {
+         return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
+     }
++    if (model.arch == LLM_ARCH_HRM_TEXT) {
++        return std::max<uint32_t>(n_tokens * 80, 64u * model.n_tensors());
++    }
+     uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());
+     for (const auto & lora : model.loras) {
+         res += lora->get_n_nodes();
+diff --git a/src/llama-hparams.h b/src/llama-hparams.h
+index e2d051edc..812598f69 100644
+--- a/src/llama-hparams.h
++++ b/src/llama-hparams.h
+@@ -164,6 +164,12 @@ struct llama_hparams {
+     float f_embedding_scale = 0.0f;
+     float f_attention_scale = 0.0f;
+ 
++    // HRM-Text recurrence metadata. n_layer remains the expanded KV-cache slot count.
++    uint32_t n_hrm_layer_per_stack = 0;
++    uint32_t n_hrm_h_cycles        = 0;
++    uint32_t n_hrm_l_cycles        = 0;
++    bool     hrm_prefix_lm         = false;
++
+     // grok-2
+     float    f_attn_out_scale = 0.0f;
+     uint32_t attn_temp_length = 0;
+diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
+index 528e4c9c0..8a6e009c6 100644
+--- a/src/llama-model-saver.cpp
++++ b/src/llama-model-saver.cpp
+@@ -245,6 +245,10 @@ void llama_model_saver::add_kv_from_model() {
+     add_kv(LLM_KV_TOKEN_SHIFT_COUNT,                 hparams.token_shift_count);
+     add_kv(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         hparams.n_moe_layer_step);
+     // add_kv(LLM_KV_FULL_ATTENTION_INTERVAL,           ???);
++    add_kv(LLM_KV_HRM_LAYERS_PER_STACK,              hparams.n_hrm_layer_per_stack);
++    add_kv(LLM_KV_HRM_H_CYCLES,                      hparams.n_hrm_h_cycles);
++    add_kv(LLM_KV_HRM_L_CYCLES,                      hparams.n_hrm_l_cycles);
++    add_kv(LLM_KV_HRM_PREFIX_LM,                     hparams.hrm_prefix_lm);
+ 
+     add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);
+     add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
+diff --git a/src/llama-model.cpp b/src/llama-model.cpp
+index 8bf20a716..a3cc996aa 100644
+--- a/src/llama-model.cpp
++++ b/src/llama-model.cpp
+@@ -96,6 +96,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
+             return new llama_model_qwen2moe(params);
+         case LLM_ARCH_QWEN3:
+             return new llama_model_qwen3(params);
++        case LLM_ARCH_HRM_TEXT:
++            return new llama_model_hrm_text(params);
+         case LLM_ARCH_QWEN3MOE:
+             return new llama_model_qwen3moe(params);
+         case LLM_ARCH_QWEN3VL:
+@@ -2339,6 +2341,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+         case LLM_ARCH_PANGU_EMBED:
+         case LLM_ARCH_AFMOE:
+         case LLM_ARCH_QWEN3NEXT:
++        case LLM_ARCH_HRM_TEXT:
+         case LLM_ARCH_MIMO2:
+         case LLM_ARCH_STEP35:
+             return LLAMA_ROPE_TYPE_NEOX;
+diff --git a/src/models/hrm-text.cpp b/src/models/hrm-text.cpp
+new file mode 100644
+index 000000000..e0a3e9f59
+--- /dev/null
++++ b/src/models/hrm-text.cpp
+@@ -0,0 +1,183 @@
++#include "models.h"
++
++#include <cmath>
++#include <vector>
++
++void llama_model_hrm_text::load_arch_hparams(llama_model_loader & ml) {
++    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
++    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale);
++    ml.get_key(LLM_KV_HRM_LAYERS_PER_STACK,        hparams.n_hrm_layer_per_stack);
++    ml.get_key(LLM_KV_HRM_H_CYCLES,                hparams.n_hrm_h_cycles);
++    ml.get_key(LLM_KV_HRM_L_CYCLES,                hparams.n_hrm_l_cycles);
++    ml.get_key(LLM_KV_HRM_PREFIX_LM,               hparams.hrm_prefix_lm, false);
++
++    switch (hparams.n_embd) {
++        case 1536: type = LLM_TYPE_1B; break;
++        default:   type = LLM_TYPE_UNKNOWN;
++    }
++}
++
++void llama_model_hrm_text::load_arch_tensors(llama_model_loader &) {
++    LLAMA_LOAD_LOCALS;
++
++    const int64_t n_stack = hparams.n_hrm_layer_per_stack;
++    const int64_t n_cycle_slots = n_stack * (hparams.n_hrm_l_cycles + 1);
++
++    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
++    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), {n_embd, n_vocab}, 0);
++
++    hrm_z_l_init = create_tensor(tn(LLM_TENSOR_HRM_Z_L_INIT), {n_embd}, 0);
++
++    std::vector<bool> loaded_physical(2 * n_stack, false);
++
++    for (int il = 0; il < n_layer; ++il) {
++        auto & layer = layers[il];
++
++        const int64_t layer_in_stack = il % n_stack;
++        const int64_t phase = (il % n_cycle_slots) / n_stack;
++        const bool is_h_stack = phase == int64_t(hparams.n_hrm_l_cycles);
++        const int physical_bid = int((is_h_stack ? n_stack : 0) + layer_in_stack);
++
++        const int flags = loaded_physical[physical_bid] ? TENSOR_DUPLICATED : 0;
++        loaded_physical[physical_bid] = true;
++
++        create_tensor_qkv(layer, physical_bid,
++                n_embd,
++                n_embd_head_k * n_head,
++                n_embd_k_gqa,
++                n_embd_v_gqa,
++                flags);
++
++        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", physical_bid), {n_embd, n_embd_head_k * n_head}, flags);
++        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,  "weight", physical_bid), {n_embd_head_k * n_head, n_embd}, flags);
++
++        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", physical_bid), {n_embd, n_ff}, flags);
++        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", physical_bid), {n_ff, n_embd}, flags);
++        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", physical_bid), {n_embd, n_ff}, flags);
++    }
++}
++
++std::unique_ptr<llm_graph_context> llama_model_hrm_text::build_arch_graph(const llm_graph_params & params) const {
++    return std::make_unique<graph>(*this, params);
++}
++
++llama_model_hrm_text::graph::graph(const llama_model & model_, const llm_graph_params & params) : llm_graph_context(params) {
++    const auto & model = static_cast<const llama_model_hrm_text &>(model_);
++
++    GGML_ASSERT(model.tok_embd != nullptr);
++    GGML_ASSERT(model.output != nullptr);
++    GGML_ASSERT(model.hrm_z_l_init != nullptr);
++
++    const int64_t n_embd_head = hparams.n_embd_head_v();
++    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
++    GGML_ASSERT(n_embd_head == n_rot);
++
++    const int64_t n_stack = hparams.n_hrm_layer_per_stack;
++    const int64_t h_cycles = hparams.n_hrm_h_cycles;
++    const int64_t l_cycles = hparams.n_hrm_l_cycles;
++
++    ggml_tensor * inp_pos = build_inp_pos();
++    auto * inp_attn = build_attn_inp_kv();
++    ggml_tensor * inp_out_ids = build_inp_out_ids();
++
++    ggml_tensor * hidden_high = build_inp_embd(model.tok_embd);
++    ggml_tensor * hidden_low = ggml_repeat(ctx0, model.hrm_z_l_init, hidden_high);
++    cb(hidden_low, "hrm_z_l_init", -1);
++
++    const float kq_scale = 1.0f / std::sqrt(float(n_embd_head));
++
++    auto build_stack = [&](ggml_tensor * stack_inp, int slot_offset) -> ggml_tensor * {
++        ggml_tensor * stack_cur = stack_inp;
++
++        for (int layer_idx = 0; layer_idx < n_stack; ++layer_idx) {
++            const int il = slot_offset + layer_idx;
++            const auto & layer = model.layers[il];
++
++            ggml_tensor * inpSA = stack_cur;
++            ggml_tensor * cur = build_norm(stack_cur, nullptr, nullptr, LLM_NORM_RMS, il);
++            cb(cur, "attn_norm", il);
++
++            {
++                ggml_tensor * attn_inp = cur;
++                auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il);
++
++                ggml_tensor * gate = build_lora_mm(layer.wqkv_gate, attn_inp, layer.wqkv_gate_s);
++                cb(gate, "attn_gate_proj", il);
++
++                Qcur = ggml_rope_ext(
++                        ctx0, Qcur, inp_pos, nullptr,
++                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                        ext_factor, attn_factor, beta_fast, beta_slow);
++                cb(Qcur, "Qcur_rope", il);
++
++                Kcur = ggml_rope_ext(
++                        ctx0, Kcur, inp_pos, nullptr,
++                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
++                        ext_factor, attn_factor, beta_fast, beta_slow);
++                cb(Kcur, "Kcur_rope", il);
++
++                cur = build_attn(inp_attn,
++                        nullptr, nullptr, nullptr,
++                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
++                cb(cur, "attn_out", il);
++
++                gate = ggml_sigmoid(ctx0, gate);
++                cb(gate, "attn_gate_sig", il);
++
++                cur = ggml_mul(ctx0, cur, gate);
++                cb(cur, "attn_gated", il);
++
++                cur = build_lora_mm(layer.wo, cur, layer.wo_s);
++                cb(cur, "attn_o_proj", il);
++            }
++
++            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
++            cb(ffn_inp, "ffn_inp", il);
++
++            cur = build_norm(ffn_inp, nullptr, nullptr, LLM_NORM_RMS, il);
++            cb(cur, "ffn_norm", il);
++
++            cur = build_ffn(cur,
++                    layer.ffn_up,   nullptr, layer.ffn_up_s,
++                    layer.ffn_gate, nullptr, layer.ffn_gate_s,
++                    layer.ffn_down, nullptr, layer.ffn_down_s,
++                    nullptr,
++                    LLM_FFN_SILU, LLM_FFN_PAR, il);
++            cb(cur, "ffn_out", il);
++
++            cur = ggml_add(ctx0, cur, ffn_inp);
++            cur = build_cvec(cur, il);
++            cb(cur, "hrm_layer_out", il);
++
++            stack_cur = cur;
++        }
++
++        stack_cur = build_norm(stack_cur, nullptr, nullptr, LLM_NORM_RMS, slot_offset);
++        cb(stack_cur, "stack_final_norm", slot_offset);
++        return stack_cur;
++    };
++
++    for (int h = 0; h < h_cycles; ++h) {
++        for (int l = 0; l < l_cycles; ++l) {
++            const int slot_offset = int((h * (l_cycles + 1) + l) * n_stack);
++            hidden_low = build_stack(ggml_add(ctx0, hidden_low, hidden_high), slot_offset);
++        }
++
++        const int slot_offset = int((h * (l_cycles + 1) + l_cycles) * n_stack);
++        hidden_high = build_stack(ggml_add(ctx0, hidden_high, hidden_low), slot_offset);
++    }
++
++    ggml_tensor * cur = hidden_high;
++
++    if (inp_out_ids) {
++        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
++    }
++
++    res->t_embd = cur;
++
++    cur = build_lora_mm(model.output, cur, model.output_s);
++    cb(cur, "result_output", -1);
++
++    res->t_logits = cur;
++    ggml_build_forward_expand(gf, cur);
++}
+diff --git a/src/models/models.h b/src/models/models.h
+index 7e551eb96..7da6b7f7f 100644
+--- a/src/models/models.h
++++ b/src/models/models.h
+@@ -515,6 +515,20 @@ struct llama_model_qwen3 : public llama_model_base {
+     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+ };
+ 
++struct llama_model_hrm_text : public llama_model_base {
++    llama_model_hrm_text(const struct llama_model_params & params) : llama_model_base(params) {}
++    void load_arch_hparams(llama_model_loader & ml) override;
++    void load_arch_tensors(llama_model_loader & ml) override;
++
++    ggml_tensor * hrm_z_l_init = nullptr;
++
++    struct graph : public llm_graph_context {
++        graph(const llama_model & model, const llm_graph_params & params);
++    };
++
++    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
++};
++
+ 
+ struct llama_model_qwen3moe : public llama_model_base {
+     llama_model_qwen3moe(const struct llama_model_params & params) : llama_model_base(params) {}