Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ 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
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,126 @@
+# Devstral-24B-NVFP4-NVembed
+
+NVFP4-quantized [Devstral-Small-2-24B-Instruct](https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct) — a 24B-parameter dense coding model that fits in 12.4 GB VRAM and runs **31K context** on a single RTX 5080.
+
+## What's special
+
+- **Full NVFP4**: all Linear layers, `lm_head`, and `embed_tokens` quantized to NVFP4
+- **FP8 KV cache**: GQA attention (32 Q / 8 KV heads), stored in FP8
+- **31,200-token context** on 16 GB VRAM (vs ~20K with BF16 embed)
+- **12.4 GiB** single-file `model.safetensors`
+- YARN rope scaling (factor=48) for long-context support
+- Standard `MistralForCausalLM` architecture — no `--trust-remote-code` needed
+
+## Hardware
+
+- **GPU**: NVIDIA RTX 5080 (16 GB VRAM, SM 12.0 / Blackwell)
+- **NVFP4 GEMM**: Marlin backend (FlashInfer FP4 JIT fails on SM 12.0)
+
+Should work on any Blackwell GPU (RTX 50-series, B-series datacenter). Context length scales with available VRAM.
+
+## Quick start
+
+### 1. Clone and build vLLM
+
+```bash
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+git checkout 628302114  # v0.16.1rc1.dev34
+
+# Apply patches (required for NVFP4 lm_head and embed_tokens)
+git apply ../Devstral-24B-NVFP4-NVembed/vllm_patches.diff
+
+# Build
+pip install -e . --no-build-isolation
+```
+
+### 2. Install dependencies
+
+```bash
+pip install flashinfer==0.6.4  # or latest compatible
+```
+
+### 3. Serve
+
+```bash
+export VLLM_TEST_FORCE_FP8_MARLIN=1
+export VLLM_NVFP4_GEMM_BACKEND=marlin
+export PYTORCH_ALLOC_CONF=expandable_segments:True
+export VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE=$((64 * 1024 * 1024))
+
+vllm serve ./Devstral-24B-NVFP4-NVembed \
+    --dtype bfloat16 \
+    --quantization modelopt \
+    --kv-cache-dtype fp8_e4m3 \
+    --max-model-len 31200 \
+    --no-enable-prefix-caching \
+    --max-num-seqs 1 \
+    --kv-cache-memory-bytes $((1950 * 1310720)) \
+    --num-gpu-blocks-override 1950 \
+    --tensor-parallel-size 1 \
+    --enable-chunked-prefill \
+    --max-num-batched-tokens 256 \
+    --enforce-eager \
+    --override-generation-config '{"temperature": 0.0, "max_tokens": 8000}'
+```
+
+### 4. Test
+
+```bash
+curl -s http://localhost:8000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Devstral-24B-NVFP4-NVembed",
+       "messages": [{"role": "user", "content": "Write a Python function to check if a number is prime."}]}' \
+  | python3 -m json.tool
+```
+
+## Serve command flags explained
+
+| Flag | Why |
+|------|-----|
+| `--quantization modelopt` | NVFP4 weight format from NVIDIA Model Optimizer |
+| `--kv-cache-dtype fp8_e4m3` | FP8 KV cache — halves KV memory vs BF16 |
+| `--no-enable-prefix-caching` | Disabled for maximum VRAM headroom |
+| `--max-num-seqs 1` | Single sequence — 16 GB VRAM budget |
+| `--kv-cache-memory-bytes` | Precise KV allocation (bypasses `gpu_memory_utilization` startup check) |
+| `--num-gpu-blocks-override 1950` | Binary-searched max stable blocks for RTX 5080 |
+| `--enforce-eager` | CUDA graphs disabled — saves VRAM |
+| `--override-generation-config` | `temperature=0` for deterministic coding output |
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Model size | 12.4 GiB |
+| Max context | 31,200 tokens |
+| KV cache blocks | 1,950 |
+| KV cache size | ~2.4 GiB |
+| KV per token | 80 KB (40 layers x 8 KV heads x 128 dim x 2 x 1 byte) |
+| Architecture | Dense, 40 layers, GQA (32 Q / 8 KV heads) |
+| Parameters | 24B |
+
+## Quantization details
+
+- **Method**: NVIDIA Model Optimizer (`nvidia-modelopt==0.41.0`)
+- **Format**: NVFP4 (4-bit floating point, group_size=16)
+- **Calibration**: 512 samples across 4 task types (code, instruction, agentic, structured data from databricks-dolly-15k)
+- **lm_head + embed_tokens**: also NVFP4 (saves ~1.9 GB combined)
+- **Source model**: Dequantized from the original FP8 checkpoint to BF16, then quantized to NVFP4
+- **Base model**: [mistralai/Devstral-Small-2-24B-Instruct](https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct)
+
+## vLLM patches
+
+The included `vllm_patches.diff` modifies 3 files based on vLLM commit `628302114`:
+
+| File | Change |
+|------|--------|
+| `modelopt.py` | NVFP4 `lm_head` + `embed_tokens` support; `PerTensorScaleParameter` init fix (uninitialized slots caused inf in merged linears) |
+| `vocab_parallel_embedding.py` | FP8 weight dtype preservation + BF16 cast in `embedding()` |
+| `linear.py` | FP8 weight cast in forward pass |
+
+### Upstream PRs
+
+Some of these patches have been submitted upstream and may become unnecessary in future vLLM releases:
+
+- [#35576](https://github.com/vllm-project/vllm/pull/35576) — MLA weight access crash fix for quantized layers
+- [#35660](https://github.com/vllm-project/vllm/pull/35660) — NVFP4-quantized lm_head support

chat_template.jinja

@@ -0,0 +1,121 @@
+{#- Default system message if no system prompt is passed. #}
+{%- set default_system_message = '' %}
+
+{#- Begin of sequence token. #}
+{{- bos_token }}
+
+{#- Handle system prompt if it exists. #}
+{#- System prompt supports text content or text chunks. #}
+{%- if messages[0]['role'] == 'system' %}
+    {{- '[SYSTEM_PROMPT]' -}}
+    {%- if messages[0]['content'] is string %}
+        {{- messages[0]['content'] -}}
+    {%- else %}        
+        {%- for block in messages[0]['content'] %}
+            {%- if block['type'] == 'text' %}
+                {{- block['text'] }}
+            {%- else %}
+                {{- raise_exception('Only text chunks are supported in system message contents.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+    {{- '[/SYSTEM_PROMPT]' -}}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set loop_messages = messages %}
+    {%- if default_system_message != '' %}
+        {{- '[SYSTEM_PROMPT]' + default_system_message + '[/SYSTEM_PROMPT]' }}
+    {%- endif %}
+{%- endif %}
+
+
+{#- Tools definition #}
+{%- set tools_definition = '' %}
+{%- set has_tools = false %}
+{%- if tools is defined and tools is not none and tools|length > 0 %}
+    {%- set has_tools = true %}
+    {%- set tools_definition = '[AVAILABLE_TOOLS]' + (tools| tojson) + '[/AVAILABLE_TOOLS]' %}
+    {{- tools_definition }}
+{%- endif %}
+
+{#- Checks for alternating user/assistant messages. #}
+{%- set ns = namespace(index=0) %}
+{%- for message in loop_messages %}
+    {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}
+        {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}
+            {{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}
+        {%- endif %}
+        {%- set ns.index = ns.index + 1 %}
+    {%- endif %}
+{%- endfor %}
+
+{#- Handle conversation messages. #}
+{%- for message in loop_messages %}
+
+    {#- User messages supports text content or text and image chunks. #}
+    {%- if message['role'] == 'user' %}
+        {%- if message['content'] is string %}
+            {{- '[INST]' + message['content'] + '[/INST]' }}
+        {%- elif message['content'] | length > 0 %}
+            {{- '[INST]' }}
+            {%- if message['content'] | length == 2 %}
+                {%- set blocks = message['content'] | sort(attribute='type') %}
+            {%- else %}
+                {%- set blocks = message['content'] %}
+            {%- endif %}
+            {%- for block in blocks %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- elif block['type'] in ['image', 'image_url'] %}
+                    {{- '[IMG]' }}
+                {%- else %}
+                    {{- raise_exception('Only text, image and image_url chunks are supported in user message content.') }}
+                {%- endif %}
+            {%- endfor %}
+            {{- '[/INST]' }}
+        {%- else %}
+            {{- raise_exception('User message must have a string or a list of chunks in content') }}
+        {%- endif %}
+
+    {#- Assistant messages supports text content or text and image chunks. #}
+    {%- elif message['role'] == 'assistant' %}
+        {%- if (message['content'] is none or message['content'] == '' or message['content']|length == 0) and (message['tool_calls'] is not defined or message['tool_calls'] is none or message['tool_calls']|length == 0) %}
+            {{- raise_exception('Assistant message must have a string or a list of chunks in content or a list of tool calls.') }}
+        {%- endif %}
+
+        {%- if message['content'] is string %}
+            {{- message['content'] }}
+        {%- elif message['content'] | length > 0 %}
+            {%- for block in message['content'] %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- else %}
+                    {{- raise_exception('Only text chunks are supported in assistant message contents.') }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        
+        {%- if message['tool_calls'] is defined and message['tool_calls'] is not none and message['tool_calls']|length > 0 %}
+            {%- for tool in message['tool_calls'] %}
+                {%- set arguments = tool['function']['arguments'] %}
+                {%- if arguments is not string %}
+                    {%- set arguments = arguments|tojson|safe %}
+                {%- elif arguments == '' %}
+                    {%- set arguments = '{}' %}
+                {%- endif %}
+                {{- '[TOOL_CALLS]' + tool['function']['name'] + '[ARGS]' + arguments }}
+            {%- endfor %}
+        {%- endif %}
+
+        {#- End of sequence token for each assistant messages. #}
+        {{- eos_token }}
+
+    {#- Tool messages only supports text content. #}
+    {%- elif message['role'] == 'tool' %}
+        {{- '[TOOL_RESULTS]' + message['content']|string + '[/TOOL_RESULTS]' }}
+
+    {#- Raise exception for unsupported roles. #}
+    {%- else %}
+        {{- raise_exception('Only user, assistant and tool roles are supported, got ' + message['role'] + '.') }}
+    {%- endif %}
+{%- endfor %}

config.json

@@ -0,0 +1,79 @@
+{
+  "architectures": [
+    "MistralForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "dtype": "bfloat16",
+  "eos_token_id": 2,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 32768,
+  "max_position_embeddings": 393216,
+  "model_type": "mistral",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "beta_fast": 32.0,
+    "beta_slow": 1.0,
+    "factor": 48.0,
+    "llama_4_scaling_beta": 0.1,
+    "mscale": 1.0,
+    "mscale_all_dim": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_theta": 100000000.0,
+    "rope_type": "yarn",
+    "type": "yarn"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.2.0",
+  "use_cache": true,
+  "vocab_size": 131072,
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "input_activations": {
+          "dynamic": false,
+          "num_bits": 4,
+          "type": "float",
+          "group_size": 16
+        },
+        "weights": {
+          "dynamic": false,
+          "num_bits": 4,
+          "type": "float",
+          "group_size": 16
+        },
+        "targets": [
+          "Linear"
+        ]
+      }
+    },
+    "ignore": [],
+    "quant_algo": "NVFP4",
+    "producer": {
+      "name": "modelopt",
+      "version": "0.41.0"
+    },
+    "quant_method": "modelopt",
+    "nvfp4_embed": true
+  },
+  "rope_scaling": {
+    "beta_fast": 32.0,
+    "beta_slow": 1.0,
+    "factor": 48.0,
+    "llama_4_scaling_beta": 0.1,
+    "mscale": 1.0,
+    "mscale_all_dim": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_theta": 100000000.0,
+    "rope_type": "yarn",
+    "type": "yarn"
+  }
+}

generation_config.json

@@ -0,0 +1,5 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "5.2.0",
+  "use_cache": true
+}

hf_quant_config.json

@@ -0,0 +1,13 @@
+{
+    "producer": {
+        "name": "modelopt",
+        "version": "0.41.0"
+    },
+    "quantization": {
+        "quant_algo": "NVFP4",
+        "kv_cache_quant_algo": null,
+        "group_size": 16,
+        "exclude_modules": [],
+        "nvfp4_embed": true
+    }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c7f9d376ec578ce2934e9db65bf58508ab738fd8e01db3da4d41830cedca88c
+size 13260207288

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c05b883079bc2f8083f6478bd21f8b4b38c82bf598c3eb6a804b323ab59e5c19
+size 17077501

tokenizer_config.json

@@ -0,0 +1,1013 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "eos_token": "</s>",
+  "extra_special_tokens": [
+    "<unk>",
+    "<s>",
+    "</s>",
+    "[INST]",
+    "[/INST]",
+    "[AVAILABLE_TOOLS]",
+    "[/AVAILABLE_TOOLS]",
+    "[TOOL_RESULTS]",
+    "[/TOOL_RESULTS]",
+    "[TOOL_CALLS]",
+    "[IMG]",
+    "<pad>",
+    "[IMG_BREAK]",
+    "[IMG_END]",
+    "[PREFIX]",
+    "[MIDDLE]",
+    "[SUFFIX]",
+    "[SYSTEM_PROMPT]",
+    "[/SYSTEM_PROMPT]",
+    "[TOOL_CONTENT]",
+    "<SPECIAL_20>",
+    "<SPECIAL_21>",
+    "<SPECIAL_22>",
+    "<SPECIAL_23>",
+    "[AUDIO]",
+    "[BEGIN_AUDIO]",
+    "<SPECIAL_26>",
+    "<SPECIAL_27>",
+    "<SPECIAL_28>",
+    "<SPECIAL_29>",
+    "<SPECIAL_30>",
+    "<SPECIAL_31>",
+    "[ARGS]",
+    "[CALL_ID]",
+    "[THINK]",
+    "[/THINK]",
+    "<SPECIAL_36>",
+    "<SPECIAL_37>",
+    "<SPECIAL_38>",
+    "<SPECIAL_39>",
+    "<SPECIAL_40>",
+    "<SPECIAL_41>",
+    "<SPECIAL_42>",
+    "<SPECIAL_43>",
+    "<SPECIAL_44>",
+    "<SPECIAL_45>",
+    "<SPECIAL_46>",
+    "<SPECIAL_47>",
+    "<SPECIAL_48>",
+    "<SPECIAL_49>",
+    "<SPECIAL_50>",
+    "<SPECIAL_51>",
+    "<SPECIAL_52>",
+    "<SPECIAL_53>",
+    "<SPECIAL_54>",
+    "<SPECIAL_55>",
+    "<SPECIAL_56>",
+    "<SPECIAL_57>",
+    "<SPECIAL_58>",
+    "<SPECIAL_59>",
+    "<SPECIAL_60>",
+    "<SPECIAL_61>",
+    "<SPECIAL_62>",
+    "<SPECIAL_63>",
+    "<SPECIAL_64>",
+    "<SPECIAL_65>",
+    "<SPECIAL_66>",
+    "<SPECIAL_67>",
+    "<SPECIAL_68>",
+    "<SPECIAL_69>",
+    "<SPECIAL_70>",
+    "<SPECIAL_71>",
+    "<SPECIAL_72>",
+    "<SPECIAL_73>",
+    "<SPECIAL_74>",
+    "<SPECIAL_75>",
+    "<SPECIAL_76>",
+    "<SPECIAL_77>",
+    "<SPECIAL_78>",
+    "<SPECIAL_79>",
+    "<SPECIAL_80>",
+    "<SPECIAL_81>",
+    "<SPECIAL_82>",
+    "<SPECIAL_83>",
+    "<SPECIAL_84>",
+    "<SPECIAL_85>",
+    "<SPECIAL_86>",
+    "<SPECIAL_87>",
+    "<SPECIAL_88>",
+    "<SPECIAL_89>",
+    "<SPECIAL_90>",
+    "<SPECIAL_91>",
+    "<SPECIAL_92>",
+    "<SPECIAL_93>",
+    "<SPECIAL_94>",
+    "<SPECIAL_95>",
+    "<SPECIAL_96>",
+    "<SPECIAL_97>",
+    "<SPECIAL_98>",
+    "<SPECIAL_99>",
+    "<SPECIAL_100>",
+    "<SPECIAL_101>",
+    "<SPECIAL_102>",
+    "<SPECIAL_103>",
+    "<SPECIAL_104>",
+    "<SPECIAL_105>",
+    "<SPECIAL_106>",
+    "<SPECIAL_107>",
+    "<SPECIAL_108>",
+    "<SPECIAL_109>",
+    "<SPECIAL_110>",
+    "<SPECIAL_111>",
+    "<SPECIAL_112>",
+    "<SPECIAL_113>",
+    "<SPECIAL_114>",
+    "<SPECIAL_115>",
+    "<SPECIAL_116>",
+    "<SPECIAL_117>",
+    "<SPECIAL_118>",
+    "<SPECIAL_119>",
+    "<SPECIAL_120>",
+    "<SPECIAL_121>",
+    "<SPECIAL_122>",
+    "<SPECIAL_123>",
+    "<SPECIAL_124>",
+    "<SPECIAL_125>",
+    "<SPECIAL_126>",
+    "<SPECIAL_127>",
+    "<SPECIAL_128>",
+    "<SPECIAL_129>",
+    "<SPECIAL_130>",
+    "<SPECIAL_131>",
+    "<SPECIAL_132>",
+    "<SPECIAL_133>",
+    "<SPECIAL_134>",
+    "<SPECIAL_135>",
+    "<SPECIAL_136>",
+    "<SPECIAL_137>",
+    "<SPECIAL_138>",
+    "<SPECIAL_139>",
+    "<SPECIAL_140>",
+    "<SPECIAL_141>",
+    "<SPECIAL_142>",
+    "<SPECIAL_143>",
+    "<SPECIAL_144>",
+    "<SPECIAL_145>",
+    "<SPECIAL_146>",
+    "<SPECIAL_147>",
+    "<SPECIAL_148>",
+    "<SPECIAL_149>",
+    "<SPECIAL_150>",
+    "<SPECIAL_151>",
+    "<SPECIAL_152>",
+    "<SPECIAL_153>",
+    "<SPECIAL_154>",
+    "<SPECIAL_155>",
+    "<SPECIAL_156>",
+    "<SPECIAL_157>",
+    "<SPECIAL_158>",
+    "<SPECIAL_159>",
+    "<SPECIAL_160>",
+    "<SPECIAL_161>",
+    "<SPECIAL_162>",
+    "<SPECIAL_163>",
+    "<SPECIAL_164>",
+    "<SPECIAL_165>",
+    "<SPECIAL_166>",
+    "<SPECIAL_167>",
+    "<SPECIAL_168>",
+    "<SPECIAL_169>",
+    "<SPECIAL_170>",
+    "<SPECIAL_171>",
+    "<SPECIAL_172>",
+    "<SPECIAL_173>",
+    "<SPECIAL_174>",
+    "<SPECIAL_175>",
+    "<SPECIAL_176>",
+    "<SPECIAL_177>",
+    "<SPECIAL_178>",
+    "<SPECIAL_179>",
+    "<SPECIAL_180>",
+    "<SPECIAL_181>",
+    "<SPECIAL_182>",
+    "<SPECIAL_183>",
+    "<SPECIAL_184>",
+    "<SPECIAL_185>",
+    "<SPECIAL_186>",
+    "<SPECIAL_187>",
+    "<SPECIAL_188>",
+    "<SPECIAL_189>",
+    "<SPECIAL_190>",
+    "<SPECIAL_191>",
+    "<SPECIAL_192>",
+    "<SPECIAL_193>",
+    "<SPECIAL_194>",
+    "<SPECIAL_195>",
+    "<SPECIAL_196>",
+    "<SPECIAL_197>",
+    "<SPECIAL_198>",
+    "<SPECIAL_199>",
+    "<SPECIAL_200>",
+    "<SPECIAL_201>",
+    "<SPECIAL_202>",
+    "<SPECIAL_203>",
+    "<SPECIAL_204>",
+    "<SPECIAL_205>",
+    "<SPECIAL_206>",
+    "<SPECIAL_207>",
+    "<SPECIAL_208>",
+    "<SPECIAL_209>",
+    "<SPECIAL_210>",
+    "<SPECIAL_211>",
+    "<SPECIAL_212>",
+    "<SPECIAL_213>",
+    "<SPECIAL_214>",
+    "<SPECIAL_215>",
+    "<SPECIAL_216>",
+    "<SPECIAL_217>",
+    "<SPECIAL_218>",
+    "<SPECIAL_219>",
+    "<SPECIAL_220>",
+    "<SPECIAL_221>",
+    "<SPECIAL_222>",
+    "<SPECIAL_223>",
+    "<SPECIAL_224>",
+    "<SPECIAL_225>",
+    "<SPECIAL_226>",
+    "<SPECIAL_227>",
+    "<SPECIAL_228>",
+    "<SPECIAL_229>",
+    "<SPECIAL_230>",
+    "<SPECIAL_231>",
+    "<SPECIAL_232>",
+    "<SPECIAL_233>",
+    "<SPECIAL_234>",
+    "<SPECIAL_235>",
+    "<SPECIAL_236>",
+    "<SPECIAL_237>",
+    "<SPECIAL_238>",
+    "<SPECIAL_239>",
+    "<SPECIAL_240>",
+    "<SPECIAL_241>",
+    "<SPECIAL_242>",
+    "<SPECIAL_243>",
+    "<SPECIAL_244>",
+    "<SPECIAL_245>",
+    "<SPECIAL_246>",
+    "<SPECIAL_247>",
+    "<SPECIAL_248>",
+    "<SPECIAL_249>",
+    "<SPECIAL_250>",
+    "<SPECIAL_251>",
+    "<SPECIAL_252>",
+    "<SPECIAL_253>",
+    "<SPECIAL_254>",
+    "<SPECIAL_255>",
+    "<SPECIAL_256>",
+    "<SPECIAL_257>",
+    "<SPECIAL_258>",
+    "<SPECIAL_259>",
+    "<SPECIAL_260>",
+    "<SPECIAL_261>",
+    "<SPECIAL_262>",
+    "<SPECIAL_263>",
+    "<SPECIAL_264>",
+    "<SPECIAL_265>",
+    "<SPECIAL_266>",
+    "<SPECIAL_267>",
+    "<SPECIAL_268>",
+    "<SPECIAL_269>",
+    "<SPECIAL_270>",
+    "<SPECIAL_271>",
+    "<SPECIAL_272>",
+    "<SPECIAL_273>",
+    "<SPECIAL_274>",
+    "<SPECIAL_275>",
+    "<SPECIAL_276>",
+    "<SPECIAL_277>",
+    "<SPECIAL_278>",
+    "<SPECIAL_279>",
+    "<SPECIAL_280>",
+    "<SPECIAL_281>",
+    "<SPECIAL_282>",
+    "<SPECIAL_283>",
+    "<SPECIAL_284>",
+    "<SPECIAL_285>",
+    "<SPECIAL_286>",
+    "<SPECIAL_287>",
+    "<SPECIAL_288>",
+    "<SPECIAL_289>",
+    "<SPECIAL_290>",
+    "<SPECIAL_291>",
+    "<SPECIAL_292>",
+    "<SPECIAL_293>",
+    "<SPECIAL_294>",
+    "<SPECIAL_295>",
+    "<SPECIAL_296>",
+    "<SPECIAL_297>",
+    "<SPECIAL_298>",
+    "<SPECIAL_299>",
+    "<SPECIAL_300>",
+    "<SPECIAL_301>",
+    "<SPECIAL_302>",
+    "<SPECIAL_303>",
+    "<SPECIAL_304>",
+    "<SPECIAL_305>",
+    "<SPECIAL_306>",
+    "<SPECIAL_307>",
+    "<SPECIAL_308>",
+    "<SPECIAL_309>",
+    "<SPECIAL_310>",
+    "<SPECIAL_311>",
+    "<SPECIAL_312>",
+    "<SPECIAL_313>",
+    "<SPECIAL_314>",
+    "<SPECIAL_315>",
+    "<SPECIAL_316>",
+    "<SPECIAL_317>",
+    "<SPECIAL_318>",
+    "<SPECIAL_319>",
+    "<SPECIAL_320>",
+    "<SPECIAL_321>",
+    "<SPECIAL_322>",
+    "<SPECIAL_323>",
+    "<SPECIAL_324>",
+    "<SPECIAL_325>",
+    "<SPECIAL_326>",
+    "<SPECIAL_327>",
+    "<SPECIAL_328>",
+    "<SPECIAL_329>",
+    "<SPECIAL_330>",
+    "<SPECIAL_331>",
+    "<SPECIAL_332>",
+    "<SPECIAL_333>",
+    "<SPECIAL_334>",
+    "<SPECIAL_335>",
+    "<SPECIAL_336>",
+    "<SPECIAL_337>",
+    "<SPECIAL_338>",
+    "<SPECIAL_339>",
+    "<SPECIAL_340>",
+    "<SPECIAL_341>",
+    "<SPECIAL_342>",
+    "<SPECIAL_343>",
+    "<SPECIAL_344>",
+    "<SPECIAL_345>",
+    "<SPECIAL_346>",
+    "<SPECIAL_347>",
+    "<SPECIAL_348>",
+    "<SPECIAL_349>",
+    "<SPECIAL_350>",
+    "<SPECIAL_351>",
+    "<SPECIAL_352>",
+    "<SPECIAL_353>",
+    "<SPECIAL_354>",
+    "<SPECIAL_355>",
+    "<SPECIAL_356>",
+    "<SPECIAL_357>",
+    "<SPECIAL_358>",
+    "<SPECIAL_359>",
+    "<SPECIAL_360>",
+    "<SPECIAL_361>",
+    "<SPECIAL_362>",
+    "<SPECIAL_363>",
+    "<SPECIAL_364>",
+    "<SPECIAL_365>",
+    "<SPECIAL_366>",
+    "<SPECIAL_367>",
+    "<SPECIAL_368>",
+    "<SPECIAL_369>",
+    "<SPECIAL_370>",
+    "<SPECIAL_371>",
+    "<SPECIAL_372>",
+    "<SPECIAL_373>",
+    "<SPECIAL_374>",
+    "<SPECIAL_375>",
+    "<SPECIAL_376>",
+    "<SPECIAL_377>",
+    "<SPECIAL_378>",
+    "<SPECIAL_379>",
+    "<SPECIAL_380>",
+    "<SPECIAL_381>",
+    "<SPECIAL_382>",
+    "<SPECIAL_383>",
+    "<SPECIAL_384>",
+    "<SPECIAL_385>",
+    "<SPECIAL_386>",
+    "<SPECIAL_387>",
+    "<SPECIAL_388>",
+    "<SPECIAL_389>",
+    "<SPECIAL_390>",
+    "<SPECIAL_391>",
+    "<SPECIAL_392>",
+    "<SPECIAL_393>",
+    "<SPECIAL_394>",
+    "<SPECIAL_395>",
+    "<SPECIAL_396>",
+    "<SPECIAL_397>",
+    "<SPECIAL_398>",
+    "<SPECIAL_399>",
+    "<SPECIAL_400>",
+    "<SPECIAL_401>",
+    "<SPECIAL_402>",
+    "<SPECIAL_403>",
+    "<SPECIAL_404>",
+    "<SPECIAL_405>",
+    "<SPECIAL_406>",
+    "<SPECIAL_407>",
+    "<SPECIAL_408>",
+    "<SPECIAL_409>",
+    "<SPECIAL_410>",
+    "<SPECIAL_411>",
+    "<SPECIAL_412>",
+    "<SPECIAL_413>",
+    "<SPECIAL_414>",
+    "<SPECIAL_415>",
+    "<SPECIAL_416>",
+    "<SPECIAL_417>",
+    "<SPECIAL_418>",
+    "<SPECIAL_419>",
+    "<SPECIAL_420>",
+    "<SPECIAL_421>",
+    "<SPECIAL_422>",
+    "<SPECIAL_423>",
+    "<SPECIAL_424>",
+    "<SPECIAL_425>",
+    "<SPECIAL_426>",
+    "<SPECIAL_427>",
+    "<SPECIAL_428>",
+    "<SPECIAL_429>",
+    "<SPECIAL_430>",
+    "<SPECIAL_431>",
+    "<SPECIAL_432>",
+    "<SPECIAL_433>",
+    "<SPECIAL_434>",
+    "<SPECIAL_435>",
+    "<SPECIAL_436>",
+    "<SPECIAL_437>",
+    "<SPECIAL_438>",
+    "<SPECIAL_439>",
+    "<SPECIAL_440>",
+    "<SPECIAL_441>",
+    "<SPECIAL_442>",
+    "<SPECIAL_443>",
+    "<SPECIAL_444>",
+    "<SPECIAL_445>",
+    "<SPECIAL_446>",
+    "<SPECIAL_447>",
+    "<SPECIAL_448>",
+    "<SPECIAL_449>",
+    "<SPECIAL_450>",
+    "<SPECIAL_451>",
+    "<SPECIAL_452>",
+    "<SPECIAL_453>",
+    "<SPECIAL_454>",
+    "<SPECIAL_455>",
+    "<SPECIAL_456>",
+    "<SPECIAL_457>",
+    "<SPECIAL_458>",
+    "<SPECIAL_459>",
+    "<SPECIAL_460>",
+    "<SPECIAL_461>",
+    "<SPECIAL_462>",
+    "<SPECIAL_463>",
+    "<SPECIAL_464>",
+    "<SPECIAL_465>",
+    "<SPECIAL_466>",
+    "<SPECIAL_467>",
+    "<SPECIAL_468>",
+    "<SPECIAL_469>",
+    "<SPECIAL_470>",
+    "<SPECIAL_471>",
+    "<SPECIAL_472>",
+    "<SPECIAL_473>",
+    "<SPECIAL_474>",
+    "<SPECIAL_475>",
+    "<SPECIAL_476>",
+    "<SPECIAL_477>",
+    "<SPECIAL_478>",
+    "<SPECIAL_479>",
+    "<SPECIAL_480>",
+    "<SPECIAL_481>",
+    "<SPECIAL_482>",
+    "<SPECIAL_483>",
+    "<SPECIAL_484>",
+    "<SPECIAL_485>",
+    "<SPECIAL_486>",
+    "<SPECIAL_487>",
+    "<SPECIAL_488>",
+    "<SPECIAL_489>",
+    "<SPECIAL_490>",
+    "<SPECIAL_491>",
+    "<SPECIAL_492>",
+    "<SPECIAL_493>",
+    "<SPECIAL_494>",
+    "<SPECIAL_495>",
+    "<SPECIAL_496>",
+    "<SPECIAL_497>",
+    "<SPECIAL_498>",
+    "<SPECIAL_499>",
+    "<SPECIAL_500>",
+    "<SPECIAL_501>",
+    "<SPECIAL_502>",
+    "<SPECIAL_503>",
+    "<SPECIAL_504>",
+    "<SPECIAL_505>",
+    "<SPECIAL_506>",
+    "<SPECIAL_507>",
+    "<SPECIAL_508>",
+    "<SPECIAL_509>",
+    "<SPECIAL_510>",
+    "<SPECIAL_511>",
+    "<SPECIAL_512>",
+    "<SPECIAL_513>",
+    "<SPECIAL_514>",
+    "<SPECIAL_515>",
+    "<SPECIAL_516>",
+    "<SPECIAL_517>",
+    "<SPECIAL_518>",
+    "<SPECIAL_519>",
+    "<SPECIAL_520>",
+    "<SPECIAL_521>",
+    "<SPECIAL_522>",
+    "<SPECIAL_523>",
+    "<SPECIAL_524>",
+    "<SPECIAL_525>",
+    "<SPECIAL_526>",
+    "<SPECIAL_527>",
+    "<SPECIAL_528>",
+    "<SPECIAL_529>",
+    "<SPECIAL_530>",
+    "<SPECIAL_531>",
+    "<SPECIAL_532>",
+    "<SPECIAL_533>",
+    "<SPECIAL_534>",
+    "<SPECIAL_535>",
+    "<SPECIAL_536>",
+    "<SPECIAL_537>",
+    "<SPECIAL_538>",
+    "<SPECIAL_539>",
+    "<SPECIAL_540>",
+    "<SPECIAL_541>",
+    "<SPECIAL_542>",
+    "<SPECIAL_543>",
+    "<SPECIAL_544>",
+    "<SPECIAL_545>",
+    "<SPECIAL_546>",
+    "<SPECIAL_547>",
+    "<SPECIAL_548>",
+    "<SPECIAL_549>",
+    "<SPECIAL_550>",
+    "<SPECIAL_551>",
+    "<SPECIAL_552>",
+    "<SPECIAL_553>",
+    "<SPECIAL_554>",
+    "<SPECIAL_555>",
+    "<SPECIAL_556>",
+    "<SPECIAL_557>",
+    "<SPECIAL_558>",
+    "<SPECIAL_559>",
+    "<SPECIAL_560>",
+    "<SPECIAL_561>",
+    "<SPECIAL_562>",
+    "<SPECIAL_563>",
+    "<SPECIAL_564>",
+    "<SPECIAL_565>",
+    "<SPECIAL_566>",
+    "<SPECIAL_567>",
+    "<SPECIAL_568>",
+    "<SPECIAL_569>",
+    "<SPECIAL_570>",
+    "<SPECIAL_571>",
+    "<SPECIAL_572>",
+    "<SPECIAL_573>",
+    "<SPECIAL_574>",
+    "<SPECIAL_575>",
+    "<SPECIAL_576>",
+    "<SPECIAL_577>",
+    "<SPECIAL_578>",
+    "<SPECIAL_579>",
+    "<SPECIAL_580>",
+    "<SPECIAL_581>",
+    "<SPECIAL_582>",
+    "<SPECIAL_583>",
+    "<SPECIAL_584>",
+    "<SPECIAL_585>",
+    "<SPECIAL_586>",
+    "<SPECIAL_587>",
+    "<SPECIAL_588>",
+    "<SPECIAL_589>",
+    "<SPECIAL_590>",
+    "<SPECIAL_591>",
+    "<SPECIAL_592>",
+    "<SPECIAL_593>",
+    "<SPECIAL_594>",
+    "<SPECIAL_595>",
+    "<SPECIAL_596>",
+    "<SPECIAL_597>",
+    "<SPECIAL_598>",
+    "<SPECIAL_599>",
+    "<SPECIAL_600>",
+    "<SPECIAL_601>",
+    "<SPECIAL_602>",
+    "<SPECIAL_603>",
+    "<SPECIAL_604>",
+    "<SPECIAL_605>",
+    "<SPECIAL_606>",
+    "<SPECIAL_607>",
+    "<SPECIAL_608>",
+    "<SPECIAL_609>",
+    "<SPECIAL_610>",
+    "<SPECIAL_611>",
+    "<SPECIAL_612>",
+    "<SPECIAL_613>",
+    "<SPECIAL_614>",
+    "<SPECIAL_615>",
+    "<SPECIAL_616>",
+    "<SPECIAL_617>",
+    "<SPECIAL_618>",
+    "<SPECIAL_619>",
+    "<SPECIAL_620>",
+    "<SPECIAL_621>",
+    "<SPECIAL_622>",
+    "<SPECIAL_623>",
+    "<SPECIAL_624>",
+    "<SPECIAL_625>",
+    "<SPECIAL_626>",
+    "<SPECIAL_627>",
+    "<SPECIAL_628>",
+    "<SPECIAL_629>",
+    "<SPECIAL_630>",
+    "<SPECIAL_631>",
+    "<SPECIAL_632>",
+    "<SPECIAL_633>",
+    "<SPECIAL_634>",
+    "<SPECIAL_635>",
+    "<SPECIAL_636>",
+    "<SPECIAL_637>",
+    "<SPECIAL_638>",
+    "<SPECIAL_639>",
+    "<SPECIAL_640>",
+    "<SPECIAL_641>",
+    "<SPECIAL_642>",
+    "<SPECIAL_643>",
+    "<SPECIAL_644>",
+    "<SPECIAL_645>",
+    "<SPECIAL_646>",
+    "<SPECIAL_647>",
+    "<SPECIAL_648>",
+    "<SPECIAL_649>",
+    "<SPECIAL_650>",
+    "<SPECIAL_651>",
+    "<SPECIAL_652>",
+    "<SPECIAL_653>",
+    "<SPECIAL_654>",
+    "<SPECIAL_655>",
+    "<SPECIAL_656>",
+    "<SPECIAL_657>",
+    "<SPECIAL_658>",
+    "<SPECIAL_659>",
+    "<SPECIAL_660>",
+    "<SPECIAL_661>",
+    "<SPECIAL_662>",
+    "<SPECIAL_663>",
+    "<SPECIAL_664>",
+    "<SPECIAL_665>",
+    "<SPECIAL_666>",
+    "<SPECIAL_667>",
+    "<SPECIAL_668>",
+    "<SPECIAL_669>",
+    "<SPECIAL_670>",
+    "<SPECIAL_671>",
+    "<SPECIAL_672>",
+    "<SPECIAL_673>",
+    "<SPECIAL_674>",
+    "<SPECIAL_675>",
+    "<SPECIAL_676>",
+    "<SPECIAL_677>",
+    "<SPECIAL_678>",
+    "<SPECIAL_679>",
+    "<SPECIAL_680>",
+    "<SPECIAL_681>",
+    "<SPECIAL_682>",
+    "<SPECIAL_683>",
+    "<SPECIAL_684>",
+    "<SPECIAL_685>",
+    "<SPECIAL_686>",
+    "<SPECIAL_687>",
+    "<SPECIAL_688>",
+    "<SPECIAL_689>",
+    "<SPECIAL_690>",
+    "<SPECIAL_691>",
+    "<SPECIAL_692>",
+    "<SPECIAL_693>",
+    "<SPECIAL_694>",
+    "<SPECIAL_695>",
+    "<SPECIAL_696>",
+    "<SPECIAL_697>",
+    "<SPECIAL_698>",
+    "<SPECIAL_699>",
+    "<SPECIAL_700>",
+    "<SPECIAL_701>",
+    "<SPECIAL_702>",
+    "<SPECIAL_703>",
+    "<SPECIAL_704>",
+    "<SPECIAL_705>",
+    "<SPECIAL_706>",
+    "<SPECIAL_707>",
+    "<SPECIAL_708>",
+    "<SPECIAL_709>",
+    "<SPECIAL_710>",
+    "<SPECIAL_711>",
+    "<SPECIAL_712>",
+    "<SPECIAL_713>",
+    "<SPECIAL_714>",
+    "<SPECIAL_715>",
+    "<SPECIAL_716>",
+    "<SPECIAL_717>",
+    "<SPECIAL_718>",
+    "<SPECIAL_719>",
+    "<SPECIAL_720>",
+    "<SPECIAL_721>",
+    "<SPECIAL_722>",
+    "<SPECIAL_723>",
+    "<SPECIAL_724>",
+    "<SPECIAL_725>",
+    "<SPECIAL_726>",
+    "<SPECIAL_727>",
+    "<SPECIAL_728>",
+    "<SPECIAL_729>",
+    "<SPECIAL_730>",
+    "<SPECIAL_731>",
+    "<SPECIAL_732>",
+    "<SPECIAL_733>",
+    "<SPECIAL_734>",
+    "<SPECIAL_735>",
+    "<SPECIAL_736>",
+    "<SPECIAL_737>",
+    "<SPECIAL_738>",
+    "<SPECIAL_739>",
+    "<SPECIAL_740>",
+    "<SPECIAL_741>",
+    "<SPECIAL_742>",
+    "<SPECIAL_743>",
+    "<SPECIAL_744>",
+    "<SPECIAL_745>",
+    "<SPECIAL_746>",
+    "<SPECIAL_747>",
+    "<SPECIAL_748>",
+    "<SPECIAL_749>",
+    "<SPECIAL_750>",
+    "<SPECIAL_751>",
+    "<SPECIAL_752>",
+    "<SPECIAL_753>",
+    "<SPECIAL_754>",
+    "<SPECIAL_755>",
+    "<SPECIAL_756>",
+    "<SPECIAL_757>",
+    "<SPECIAL_758>",
+    "<SPECIAL_759>",
+    "<SPECIAL_760>",
+    "<SPECIAL_761>",
+    "<SPECIAL_762>",
+    "<SPECIAL_763>",
+    "<SPECIAL_764>",
+    "<SPECIAL_765>",
+    "<SPECIAL_766>",
+    "<SPECIAL_767>",
+    "<SPECIAL_768>",
+    "<SPECIAL_769>",
+    "<SPECIAL_770>",
+    "<SPECIAL_771>",
+    "<SPECIAL_772>",
+    "<SPECIAL_773>",
+    "<SPECIAL_774>",
+    "<SPECIAL_775>",
+    "<SPECIAL_776>",
+    "<SPECIAL_777>",
+    "<SPECIAL_778>",
+    "<SPECIAL_779>",
+    "<SPECIAL_780>",
+    "<SPECIAL_781>",
+    "<SPECIAL_782>",
+    "<SPECIAL_783>",
+    "<SPECIAL_784>",
+    "<SPECIAL_785>",
+    "<SPECIAL_786>",
+    "<SPECIAL_787>",
+    "<SPECIAL_788>",
+    "<SPECIAL_789>",
+    "<SPECIAL_790>",
+    "<SPECIAL_791>",
+    "<SPECIAL_792>",
+    "<SPECIAL_793>",
+    "<SPECIAL_794>",
+    "<SPECIAL_795>",
+    "<SPECIAL_796>",
+    "<SPECIAL_797>",
+    "<SPECIAL_798>",
+    "<SPECIAL_799>",
+    "<SPECIAL_800>",
+    "<SPECIAL_801>",
+    "<SPECIAL_802>",
+    "<SPECIAL_803>",
+    "<SPECIAL_804>",
+    "<SPECIAL_805>",
+    "<SPECIAL_806>",
+    "<SPECIAL_807>",
+    "<SPECIAL_808>",
+    "<SPECIAL_809>",
+    "<SPECIAL_810>",
+    "<SPECIAL_811>",
+    "<SPECIAL_812>",
+    "<SPECIAL_813>",
+    "<SPECIAL_814>",
+    "<SPECIAL_815>",
+    "<SPECIAL_816>",
+    "<SPECIAL_817>",
+    "<SPECIAL_818>",
+    "<SPECIAL_819>",
+    "<SPECIAL_820>",
+    "<SPECIAL_821>",
+    "<SPECIAL_822>",
+    "<SPECIAL_823>",
+    "<SPECIAL_824>",
+    "<SPECIAL_825>",
+    "<SPECIAL_826>",
+    "<SPECIAL_827>",
+    "<SPECIAL_828>",
+    "<SPECIAL_829>",
+    "<SPECIAL_830>",
+    "<SPECIAL_831>",
+    "<SPECIAL_832>",
+    "<SPECIAL_833>",
+    "<SPECIAL_834>",
+    "<SPECIAL_835>",
+    "<SPECIAL_836>",
+    "<SPECIAL_837>",
+    "<SPECIAL_838>",
+    "<SPECIAL_839>",
+    "<SPECIAL_840>",
+    "<SPECIAL_841>",
+    "<SPECIAL_842>",
+    "<SPECIAL_843>",
+    "<SPECIAL_844>",
+    "<SPECIAL_845>",
+    "<SPECIAL_846>",
+    "<SPECIAL_847>",
+    "<SPECIAL_848>",
+    "<SPECIAL_849>",
+    "<SPECIAL_850>",
+    "<SPECIAL_851>",
+    "<SPECIAL_852>",
+    "<SPECIAL_853>",
+    "<SPECIAL_854>",
+    "<SPECIAL_855>",
+    "<SPECIAL_856>",
+    "<SPECIAL_857>",
+    "<SPECIAL_858>",
+    "<SPECIAL_859>",
+    "<SPECIAL_860>",
+    "<SPECIAL_861>",
+    "<SPECIAL_862>",
+    "<SPECIAL_863>",
+    "<SPECIAL_864>",
+    "<SPECIAL_865>",
+    "<SPECIAL_866>",
+    "<SPECIAL_867>",
+    "<SPECIAL_868>",
+    "<SPECIAL_869>",
+    "<SPECIAL_870>",
+    "<SPECIAL_871>",
+    "<SPECIAL_872>",
+    "<SPECIAL_873>",
+    "<SPECIAL_874>",
+    "<SPECIAL_875>",
+    "<SPECIAL_876>",
+    "<SPECIAL_877>",
+    "<SPECIAL_878>",
+    "<SPECIAL_879>",
+    "<SPECIAL_880>",
+    "<SPECIAL_881>",
+    "<SPECIAL_882>",
+    "<SPECIAL_883>",
+    "<SPECIAL_884>",
+    "<SPECIAL_885>",
+    "<SPECIAL_886>",
+    "<SPECIAL_887>",
+    "<SPECIAL_888>",
+    "<SPECIAL_889>",
+    "<SPECIAL_890>",
+    "<SPECIAL_891>",
+    "<SPECIAL_892>",
+    "<SPECIAL_893>",
+    "<SPECIAL_894>",
+    "<SPECIAL_895>",
+    "<SPECIAL_896>",
+    "<SPECIAL_897>",
+    "<SPECIAL_898>",
+    "<SPECIAL_899>",
+    "<SPECIAL_900>",
+    "<SPECIAL_901>",
+    "<SPECIAL_902>",
+    "<SPECIAL_903>",
+    "<SPECIAL_904>",
+    "<SPECIAL_905>",
+    "<SPECIAL_906>",
+    "<SPECIAL_907>",
+    "<SPECIAL_908>",
+    "<SPECIAL_909>",
+    "<SPECIAL_910>",
+    "<SPECIAL_911>",
+    "<SPECIAL_912>",
+    "<SPECIAL_913>",
+    "<SPECIAL_914>",
+    "<SPECIAL_915>",
+    "<SPECIAL_916>",
+    "<SPECIAL_917>",
+    "<SPECIAL_918>",
+    "<SPECIAL_919>",
+    "<SPECIAL_920>",
+    "<SPECIAL_921>",
+    "<SPECIAL_922>",
+    "<SPECIAL_923>",
+    "<SPECIAL_924>",
+    "<SPECIAL_925>",
+    "<SPECIAL_926>",
+    "<SPECIAL_927>",
+    "<SPECIAL_928>",
+    "<SPECIAL_929>",
+    "<SPECIAL_930>",
+    "<SPECIAL_931>",
+    "<SPECIAL_932>",
+    "<SPECIAL_933>",
+    "<SPECIAL_934>",
+    "<SPECIAL_935>",
+    "<SPECIAL_936>",
+    "<SPECIAL_937>",
+    "<SPECIAL_938>",
+    "<SPECIAL_939>",
+    "<SPECIAL_940>",
+    "<SPECIAL_941>",
+    "<SPECIAL_942>",
+    "<SPECIAL_943>",
+    "<SPECIAL_944>",
+    "<SPECIAL_945>",
+    "<SPECIAL_946>",
+    "<SPECIAL_947>",
+    "<SPECIAL_948>",
+    "<SPECIAL_949>",
+    "<SPECIAL_950>",
+    "<SPECIAL_951>",
+    "<SPECIAL_952>",
+    "<SPECIAL_953>",
+    "<SPECIAL_954>",
+    "<SPECIAL_955>",
+    "<SPECIAL_956>",
+    "<SPECIAL_957>",
+    "<SPECIAL_958>",
+    "<SPECIAL_959>",
+    "<SPECIAL_960>",
+    "<SPECIAL_961>",
+    "<SPECIAL_962>",
+    "<SPECIAL_963>",
+    "<SPECIAL_964>",
+    "<SPECIAL_965>",
+    "<SPECIAL_966>",
+    "<SPECIAL_967>",
+    "<SPECIAL_968>",
+    "<SPECIAL_969>",
+    "<SPECIAL_970>",
+    "<SPECIAL_971>",
+    "<SPECIAL_972>",
+    "<SPECIAL_973>",
+    "<SPECIAL_974>",
+    "<SPECIAL_975>",
+    "<SPECIAL_976>",
+    "<SPECIAL_977>",
+    "<SPECIAL_978>",
+    "<SPECIAL_979>",
+    "<SPECIAL_980>",
+    "<SPECIAL_981>",
+    "<SPECIAL_982>",
+    "<SPECIAL_983>",
+    "<SPECIAL_984>",
+    "<SPECIAL_985>",
+    "<SPECIAL_986>",
+    "<SPECIAL_987>",
+    "<SPECIAL_988>",
+    "<SPECIAL_989>",
+    "<SPECIAL_990>",
+    "<SPECIAL_991>",
+    "<SPECIAL_992>",
+    "<SPECIAL_993>",
+    "<SPECIAL_994>",
+    "<SPECIAL_995>",
+    "<SPECIAL_996>",
+    "<SPECIAL_997>",
+    "<SPECIAL_998>",
+    "<SPECIAL_999>"
+  ],
+  "is_local": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "processor_class": "PixtralProcessor",
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<unk>"
+}

vllm_patches.diff

@@ -0,0 +1,413 @@
+diff --git a/vllm/model_executor/layers/linear.py b/vllm/model_executor/layers/linear.py
+index 5fc9fa073..ddb758886 100644
+--- a/vllm/model_executor/layers/linear.py
++++ b/vllm/model_executor/layers/linear.py
+@@ -257,13 +257,17 @@ class UnquantizedLinearMethod(LinearMethodBase):
+         x: torch.Tensor,
+         bias: torch.Tensor | None = None,
+     ) -> torch.Tensor:
++        weight = layer.weight
++        # Support FP8 weight storage: cast to compute dtype for GEMM
++        if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
++            weight = weight.to(x.dtype)
+         if (
+             vllm_is_batch_invariant()
+             and current_platform.is_cuda_alike()
+             and is_layer_moe_router_gate(getattr(layer, "prefix", ""))
+         ):
+-            return linear_batch_invariant(x, layer.weight, bias)
+-        return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
++            return linear_batch_invariant(x, weight, bias)
++        return dispatch_unquantized_gemm()(layer, x, weight, bias)
+ 
+ 
+ class LinearBase(PluggableLayer):
+diff --git a/vllm/model_executor/layers/quantization/modelopt.py b/vllm/model_executor/layers/quantization/modelopt.py
+index 999bb6325..35db5468b 100644
+--- a/vllm/model_executor/layers/quantization/modelopt.py
++++ b/vllm/model_executor/layers/quantization/modelopt.py
+@@ -43,6 +43,10 @@ from vllm.model_executor.layers.linear import (
+     LinearMethodBase,
+     UnquantizedLinearMethod,
+ )
++from vllm.model_executor.layers.vocab_parallel_embedding import (
++    ParallelLMHead,
++    VocabParallelEmbedding,
++)
+ from vllm.model_executor.layers.quantization import QuantizationMethods
+ from vllm.model_executor.layers.quantization.base_config import (
+     QuantizationConfig,
+@@ -190,6 +194,8 @@ class ModelOptQuantConfigBase(QuantizationConfig):
+         if self.is_layer_excluded(prefix):
+             if isinstance(layer, LinearBase):
+                 return UnquantizedLinearMethod()
++            if isinstance(layer, ParallelLMHead):
++                return None  # falls back to UnquantizedEmbeddingMethod
+             return None
+ 
+         # TODO: This special hard coded logic is not needed for quantized checkpoints
+@@ -197,7 +203,7 @@ class ModelOptQuantConfigBase(QuantizationConfig):
+         # exclude_modules config. But need to keep them for loading quantized
+         # checkpoints generated by older versions. Then check substring matching
+         # for patterns not caught by exact match
+-        if "vision_tower" in prefix or "vision_model" in prefix:
++        if "vision_tower" in prefix or "vision_model" in prefix or "visual" in prefix:
+             return UnquantizedLinearMethod()
+ 
+         # now, the layer is quantized, handle it here
+@@ -213,6 +219,18 @@ class ModelOptQuantConfigBase(QuantizationConfig):
+             if getattr(quant_method, "backend", "") == "marlin":
+                 quant_method.marlin_input_dtype = get_marlin_input_dtype(prefix)
+             return quant_method
++        elif isinstance(layer, ParallelLMHead):
++            # NVFP4-quantized lm_head: use LMHead-specific method that
++            # handles VocabParallelEmbedding's weight_loader.
++            quant_method = ModelOptNvFp4LMHeadMethod(self)
++            if getattr(quant_method, "backend", "") == "marlin":
++                quant_method.marlin_input_dtype = get_marlin_input_dtype(prefix)
++            return quant_method
++        elif isinstance(layer, VocabParallelEmbedding) and self.nvfp4_embed:
++            # NVFP4-quantized embed_tokens: use custom embedding method
++            # that does per-row dequantization during lookup.
++            assert isinstance(self, ModelOptNvFp4Config)
++            return ModelOptNvFp4EmbeddingMethod(self)
+ 
+         return None
+ 
+@@ -1014,9 +1032,11 @@ class ModelOptNvFp4Config(ModelOptQuantConfigBase):
+         kv_cache_quant_algo: str | None,
+         exclude_modules: list[str],
+         group_size: int = 16,
++        nvfp4_embed: bool = False,
+     ) -> None:
+         super().__init__(exclude_modules)
+         self.is_checkpoint_nvfp4_serialized = is_checkpoint_nvfp4_serialized
++        self.nvfp4_embed = nvfp4_embed
+         if is_checkpoint_nvfp4_serialized:
+             logger.warning(
+                 "Detected ModelOpt NVFP4 checkpoint. Please note that"
+@@ -1075,11 +1095,22 @@ class ModelOptNvFp4Config(ModelOptQuantConfigBase):
+                     f"hf_quant_config.json: {missing_fields}"
+                 )
+ 
++        nvfp4_embed = False
++        if "quantization" in original_config:
++            # hf_quant_config.json format: {"quantization": {"nvfp4_embed": true}}
++            nvfp4_embed = original_config["quantization"].get(
++                "nvfp4_embed", False
++            )
++        else:
++            # compressed-tensors style format (config.json quantization_config)
++            nvfp4_embed = original_config.get("nvfp4_embed", False)
++
+         return cls(
+             is_checkpoint_nvfp4_serialized,
+             kv_cache_quant_method,
+             exclude_modules,
+             group_size,
++            nvfp4_embed,
+         )
+ 
+ 
+@@ -1150,6 +1181,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
+             data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+             weight_loader=weight_loader,
+         )
++        input_global_scale[:] = torch.finfo(torch.float32).min
+         layer.register_parameter("input_scale", input_global_scale)
+ 
+         # Weight Global Scale
+@@ -1157,6 +1189,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
+             data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+             weight_loader=weight_loader,
+         )
++        weight_global_scale[:] = torch.finfo(torch.float32).min
+         layer.register_parameter("weight_scale_2", weight_global_scale)
+ 
+         # Per Block Weight Scale
+@@ -1207,6 +1240,245 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
+         )
+ 
+ 
++class ModelOptNvFp4LMHeadMethod(ModelOptNvFp4LinearMethod):
++    """NVFP4 method for ParallelLMHead (extends VocabParallelEmbedding).
++
++    ParallelLMHead passes VocabParallelEmbedding.weight_loader to
++    create_weights, but that loader doesn't handle NVFP4 packed uint8
++    weights or PerTensorScaleParameter scalars.  We override create_weights
++    to supply a compatible weight_loader that handles vocab-parallel
++    sharding for NVFP4 parameters.
++    """
++
++    def create_weights(
++        self,
++        layer: torch.nn.Module,
++        input_size_per_partition: int,
++        output_partition_sizes: list[int],
++        input_size: int,
++        output_size: int,
++        params_dtype: torch.dtype,
++        **extra_weight_attrs,
++    ):
++        # Build a weight_loader that works with NVFP4 packed parameters
++        # on a VocabParallelEmbedding-based layer.
++        embedding = layer  # layer is the ParallelLMHead instance
++
++        def _nvfp4_lm_head_weight_loader(
++            param: torch.Tensor,
++            loaded_weight: torch.Tensor,
++        ):
++            output_dim = getattr(param, "output_dim", None)
++            if output_dim is None:
++                # Scalar parameters (input_scale, weight_scale_2).
++                # Checkpoint stores a scalar; parameter is 1-element tensor.
++                if param.data.dim() == 1 and loaded_weight.dim() == 0:
++                    param.data[0] = loaded_weight.item()
++                else:
++                    param.data.copy_(loaded_weight)
++                return
++
++            # Sharded parameters (weight, weight_scale): use the
++            # VocabParallelEmbedding shard indices for the output dim.
++            tp_size = embedding.tp_size
++            if tp_size == 1:
++                param.data.copy_(loaded_weight)
++            else:
++                start = embedding.shard_indices.org_vocab_start_index
++                end = embedding.shard_indices.org_vocab_end_index
++                shard_size = end - start
++                input_dim = getattr(param, "input_dim", None)
++                # Scale start/end by the packing ratio for packed dims
++                if input_dim is not None and input_dim == output_dim:
++                    raise NotImplementedError(
++                        "Packed dim == output dim not supported for "
++                        "NVFP4 LM head"
++                    )
++                loaded_weight = loaded_weight.narrow(
++                    output_dim, start, shard_size
++                )
++                param.data[: loaded_weight.shape[0]].copy_(loaded_weight)
++                param.data[loaded_weight.shape[0] :].fill_(0)
++
++        extra_weight_attrs["weight_loader"] = _nvfp4_lm_head_weight_loader
++
++        # LinearBase sets params_dtype; VocabParallelEmbedding doesn't.
++        # Marlin's prepare_fp4_layer_for_marlin needs it for scale dtype.
++        if not hasattr(layer, "params_dtype"):
++            layer.params_dtype = params_dtype
++
++        super().create_weights(
++            layer,
++            input_size_per_partition,
++            output_partition_sizes,
++            input_size,
++            output_size,
++            params_dtype,
++            **extra_weight_attrs,
++        )
++
++
++class ModelOptNvFp4EmbeddingMethod(QuantizeMethodBase):
++    """NVFP4 method for VocabParallelEmbedding (embed_tokens).
++
++    Stores embedding weights in NVFP4 packed format to save VRAM
++    (320 MB vs 1280 MB for BF16 with 131072×5120 embeddings).
++    Dequantizes per-row during lookup.
++    """
++
++    # FP4 E2M1 lookup table: maps 4-bit index to float value
++    FP4_E2M1_LUT = [
++        0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0,
++        0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0,
++    ]
++
++    def __init__(self, quant_config: "ModelOptNvFp4Config") -> None:
++        self.quant_config = quant_config
++
++    def create_weights(
++        self,
++        layer: torch.nn.Module,
++        input_size_per_partition: int,
++        output_partition_sizes: list[int],
++        input_size: int,
++        output_size: int,
++        params_dtype: torch.dtype,
++        **extra_weight_attrs,
++    ):
++        embedding = layer  # layer is the VocabParallelEmbedding instance
++        output_size_per_partition = sum(output_partition_sizes)
++        group_size = self.quant_config.group_size
++
++        def _nvfp4_embed_weight_loader(
++            param: torch.Tensor,
++            loaded_weight: torch.Tensor,
++        ):
++            output_dim = getattr(param, "output_dim", None)
++            if output_dim is None:
++                # Scalar parameters (input_scale, weight_scale_2)
++                if param.data.dim() == 1 and loaded_weight.dim() == 0:
++                    param.data[0] = loaded_weight.item()
++                else:
++                    param.data.copy_(loaded_weight)
++                return
++
++            tp_size = embedding.tp_size
++            if tp_size == 1:
++                param.data.copy_(loaded_weight)
++            else:
++                start = embedding.shard_indices.org_vocab_start_index
++                end = embedding.shard_indices.org_vocab_end_index
++                shard_size = end - start
++                loaded_weight = loaded_weight.narrow(
++                    output_dim, start, shard_size
++                )
++                param.data[: loaded_weight.shape[0]].copy_(loaded_weight)
++                param.data[loaded_weight.shape[0] :].fill_(0)
++
++        # Packed NVFP4 weight: [vocab, hidden/2] uint8
++        from vllm.model_executor.parameter import ModelWeightParameter
++        weight = ModelWeightParameter(
++            data=torch.empty(
++                output_size_per_partition,
++                input_size_per_partition // 2,
++                dtype=torch.uint8,
++            ),
++            input_dim=1,
++            output_dim=0,
++            weight_loader=_nvfp4_embed_weight_loader,
++        )
++        layer.register_parameter("weight", weight)
++
++        # Per-group scale: [vocab, hidden/group_size] float8_e4m3fn
++        weight_scale = ModelWeightParameter(
++            data=torch.empty(
++                output_size_per_partition,
++                input_size_per_partition // group_size,
++                dtype=torch.float8_e4m3fn,
++            ),
++            input_dim=1,
++            output_dim=0,
++            weight_loader=_nvfp4_embed_weight_loader,
++        )
++        layer.register_parameter("weight_scale", weight_scale)
++
++        # Global scales (scalars)
++        from vllm.model_executor.parameter import PerTensorScaleParameter
++        input_scale = PerTensorScaleParameter(
++            data=torch.empty(1, dtype=torch.float32),
++            weight_loader=_nvfp4_embed_weight_loader,
++        )
++        layer.register_parameter("input_scale", input_scale)
++
++        weight_scale_2 = PerTensorScaleParameter(
++            data=torch.empty(1, dtype=torch.float32),
++            weight_loader=_nvfp4_embed_weight_loader,
++        )
++        layer.register_parameter("weight_scale_2", weight_scale_2)
++
++    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
++        # Merge global scales into a single alpha factor
++        input_gs = layer.input_scale.max().to(torch.float32)
++        weight_gs = layer.weight_scale_2.max().to(torch.float32)
++        layer.global_scale = Parameter(
++            input_gs * weight_gs, requires_grad=False
++        )
++        del layer.input_scale, layer.weight_scale_2
++
++        # Build FP4 E2M1 lookup table on the correct device
++        layer.fp4_lut = torch.tensor(
++            self.FP4_E2M1_LUT, dtype=torch.bfloat16,
++            device=layer.weight.device,
++        )
++
++    def embedding(
++        self, layer: torch.nn.Module, input_: torch.Tensor,
++    ) -> torch.Tensor:
++        """NVFP4 embedding lookup with per-row dequantization."""
++        group_size = self.quant_config.group_size
++        packed = layer.weight  # [vocab, hidden/2] uint8
++        scales = layer.weight_scale  # [vocab, hidden/group_size] fp8
++        global_scale = layer.global_scale  # scalar
++        fp4_lut = layer.fp4_lut  # [16] bf16
++
++        # 1. Index into packed arrays by token IDs
++        ids = input_.long()
++        packed_rows = packed[ids]  # [..., hidden/2] uint8
++        scale_rows = scales[ids]  # [..., hidden/group_size] fp8
++
++        # 2. Unpack uint8 → two FP4 indices (low nibble first)
++        low = (packed_rows & 0x0F).long()  # [..., hidden/2]
++        high = (packed_rows >> 4).long()  # [..., hidden/2]
++        # Interleave: element 2i = low[i], element 2i+1 = high[i]
++        unpacked = torch.stack([low, high], dim=-1)  # [..., hidden/2, 2]
++        orig_shape = list(ids.shape) + [packed.shape[-1] * 2]
++        unpacked = unpacked.reshape(orig_shape)  # [..., hidden]
++
++        # 3. FP4 → float via lookup table
++        dequant = fp4_lut[unpacked]  # [..., hidden] bf16
++
++        # 4. Apply per-group scale and global scale
++        scale_rows_bf16 = scale_rows.to(torch.bfloat16)  # [..., hidden/gs]
++        scale_rows_expanded = scale_rows_bf16.repeat_interleave(
++            group_size, dim=-1
++        )  # [..., hidden]
++        dequant = dequant * scale_rows_expanded * global_scale.to(
++            torch.bfloat16
++        )
++
++        return dequant
++
++    def apply(
++        self,
++        layer: torch.nn.Module,
++        x: torch.Tensor,
++        bias: torch.Tensor | None = None,
++    ) -> torch.Tensor:
++        raise NotImplementedError(
++            "NVFP4 embedding does not support linear apply"
++        )
++
++
+ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
+     """
+     MoE Method for FP4 Quantization.
+diff --git a/vllm/model_executor/layers/vocab_parallel_embedding.py b/vllm/model_executor/layers/vocab_parallel_embedding.py
+index daaa86bed..ac5bedb6b 100644
+--- a/vllm/model_executor/layers/vocab_parallel_embedding.py
++++ b/vllm/model_executor/layers/vocab_parallel_embedding.py
+@@ -66,10 +66,18 @@ class UnquantizedEmbeddingMethod(QuantizeMethodBase):
+         x: torch.Tensor,
+         bias: torch.Tensor | None = None,
+     ) -> torch.Tensor:
+-        return dispatch_unquantized_gemm()(layer, x, layer.weight, bias)
++        weight = layer.weight
++        # Support FP8 weight storage: cast to compute dtype for GEMM
++        if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
++            weight = weight.to(x.dtype)
++        return dispatch_unquantized_gemm()(layer, x, weight, bias)
+ 
+     def embedding(self, layer: torch.nn.Module, input_: torch.Tensor) -> torch.Tensor:
+-        return F.embedding(input_, layer.weight)
++        output = F.embedding(input_, layer.weight)
++        # Support FP8 weight storage: cast to compute dtype after lookup
++        if output.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
++            output = output.to(torch.bfloat16)
++        return output
+ 
+ 
+ def pad_vocab_size(vocab_size: int, pad_to: int = DEFAULT_VOCAB_PADDING_SIZE) -> int:
+@@ -429,6 +437,14 @@ class VocabParallelEmbedding(CustomOp):
+                 shape[output_dim] = self.num_embeddings_per_partition
+             param.materialize(tuple(shape), dtype=loaded_weight.dtype)
+ 
++        # If loaded weight is FP8, cast parameter to match so FP8 is
++        # preserved in memory (saves VRAM for e.g. embed_tokens).
++        if loaded_weight.dtype in (torch.float8_e4m3fn,
++                                   torch.float8_e5m2):
++            if param.data.dtype != loaded_weight.dtype:
++                param.data = torch.empty_like(param.data,
++                                              dtype=loaded_weight.dtype)
++
+         # If parameter does not have output dim, then it should
+         # be copied onto all gpus (e.g. g_idx for act_order gptq).
+         if output_dim is None: